into a separate transaction_participant structure

handlerton inherits it, so handlerton itself doesn't change.
but entities that only need to participate in a transaction,
like binlog or online alter log, use a transaction_participant
and no longer need to pretend to be a full-blown but invisible
storage engine which doesn't support create table.

remove unused methods, reorder methods, add comments

1. randomize all vectors via multiplication by a random orthogonal
   matrix
   * to generate the matrix fill the square matrix with normally
     distributed random values and create an orthogonal matrix with
     the QR decomposition
   * the rnd generator is seeded with the number of dimensions,
     so the matrix will be always the same for a given table
   * multiplication by an orthogonal matrix is a "rotation", so
     does not change distances or angles
2. when calculating the distance, first calculate a "subdistance",
   the distance between projections to the first subdist_part
   coordinates (=192, best by test, if it's larger it's less efficient,
   if it's smaller the error rate is too high)
3. calculate the full distance only if "subdistance" isn't confidently
   higher (above subdist_margin) than the distance we're comparing with
   * it might look like it would make sense to do a second projection
     at, say, subdist_part*2, and so on - but in practice one check
     is enough, the projected distance converges quickly and if it
     isn't confidently higher at subdist_part, it won't be later either

This optimization introduces a constant overhead per insert/search
operation - an input/query vector has to be multiplied by the matrix.
And the optimization saves on every distance calculation. Thus it is only
beneficial when a number of distance calculations (which grows with M
and with the table size) is high enough to outweigh the constant
overhead. Let's use MIN_ROWS table option to estimate the number of rows
in the table. use_subdist_heuristic() is optimal for mnist and
fashion-mnist (784 dimensions, 60k rows) and variations of gist (960
dimensions, 200k, 400k, 600k, 800k, 1000k rows)

Rename high-level indexes along with a table.

This patch fixes only TRUNCATE by recreate variant, there seem to be no
reasonable engine that uses TRUNCATE by handler method for testing.

Reset index_cinfo so that mi_create is not confused by garbage passed via
index_file_name and sets MY_DELETE_OLD flag.

Review question: can we add a test case to make sure VECTOR index is empty
indeed?

This commit introduces two utility functions meant to make working with
vectors simpler.

Vec_ToText converts a binary vector into a json array of numbers
(floats).
Vec_FromText takes in a json array of numbers and converts it into a
little-endian IEEE float sequence of bytes (4 bytes per float).

This method will write out a float to a String object, keeping the
charset of the original string.

Also have Float::to_string make use of String::append_float

introduced a generosity factor that makes the search less greedy.
it dramatically improves the recall by making the search a bit slower
(for the same recall one can use half the M and smaller ef).

had to add Queue::safe_push() method that removes one of the
furthest elements (not necessarily the furthest) in the queue
to keep it from overflowing.

to return only as many elements as needed, the caller no longer needs to
overallocate result arrays for throwaway nodes

use int16_t instead of floats, they're faster and smaller.
but perform intermediate SIMD calculations with floats to avoid overflows.
recall drop with such scheme is below 0.002, often none.

int8_t would've been better but the precision loss is too big
and recall degrades too much.

make handler::prepare_for_insert() to be called to prepare
the handler for writes, INSERT/UPDATE/DELETE.

When the source row is deleted, mark the corresponding node in HNSW
index by setting `tref` to null. An index is added for the `tref` in
secondary table for faster searching of the to-be-marked nodes.

The nodes marked as deleted will still be used for search, but will not
be included in the final query results.

As skipping deleted nodes and not adding deleted nodes for new-inserted
nodes' neighbor list could impact the performance, we now only skip
these nodes in search results.

- for some reason the bitmap is not set for hlindex during the delete so
  I had to temporarily comment out one line

All new code of the whole pull request, including one or several files
that are either new files or modified ones, are contributed under the
BSD-new license. I am contributing on behalf of my employer Amazon Web
Services, Inc.

* preserve the graph in memory between statements
* keep it in a TABLE_SHARE, available for concurrent searches
* nodes are generally read-only, walking the graph doesn't change them
* distance to target is cached, calculated only once
* SIMD-optimized bloom filter detects visited nodes
* nodes are stored in an array, not List, to better utilize bloom filter
* auto-adjusting heuristic to estimate the number of visited nodes
  (to configure the bloom filter)
* many threads can concurrently walk the graph. MEM_ROOT and Hash_set
  are protected with a mutex, but walking doesn't need them
* up to 8 threads can concurrently load nodes into the cache,
  nodes are partitioned into 8 mutexes (8 is chosen arbitrarily, might
  need tuning)
* concurrent editing is not supported though
* this is fine for MyISAM, TL_WRITE protects the TABLE_SHARE and the
  graph (note that TL_WRITE_CONCURRENT_INSERT is not allowed, because an
  INSERT into the main table means multiple UPDATEs in the graph)
* InnoDB uses secondary transaction-level caches linked in a list in
  in thd->ha_data via a fake handlerton
* on rollback the secondary cache is discarded, on commit nodes
  from the secondary cache are invalidated in the shared cache
  while it is exclusively locked
* on savepoint rollback both caches are flushed. this can be improved
  in the future with a row visibility callback
* graph size is controlled by @@mhnsw_cache_size, the cache is flushed
  when it reaches the threshold

instead of one row per node per layer, have one row per node.
store all neighbors for all layers in that row, and the vector itself too

it completely avoids searches in the graph table and
will allow to implement deletions in the future

1. introduce alpha. the value of 1.1 is optimal, so hard-code it.

2. hard-code ef_construction=10, best by test

3. rename hnsw_max_connection_per_layer to mhnsw_max_edges_per_node
   (max_connection is rather ambiguous in MariaDB) and add a help text

4. rename hnsw_ef_search to mhnsw_min_limit and add a help text

* mhnsw:
  * use primary key, innodb loves and (and the index cannot have dupes anyway)
    * MyISAM is ok with that, performance-wise
  * must be ha_rnd_init(0) because we aren't going to scan
    * MyISAM resets the position on ha_rnd_init(0) so query it before
    * oh, and use the correct handler, just in case
  * HA_ERR_RECORD_IS_THE_SAME is no error
* innodb:
  * return ref_length on create
  * don't assume table->pos_in_table_list is set
  * ok, assume away, but only for system versioned tables
* set alter_info on create (InnoDB needs to check for FKs)
* pair external_lock/external_unlock correctly

otherwise it'll be free'd twice

also add missing candidates.empty();

we know it every time