LPBUG 756966 is solved.

 - MWL#116 Group commit
 - MWL#136 Enhancements for START TRANSACTION WITH CONSISTENT SNAPSHOT
 - MWL#47 Annotate_rows_log_event
 - MWL#163 innodb_release_locks_early
 - Percona patch enhancing row-based replication for tables with no primary key

- Let advance_sj_state() save the value of JOIN::cur_dups_producing_tables
  in POSITION::prefix_dups_producing_tables, and restore_sj_state() restore
  it. 

Also implement the InnoDB changes for group commit into innodb_plugin.

hold the latch across the 2-phase commit, and it is already released at
start of prepare (if not before).

Rename trx->active_trans to active_flag; since we changed the semantics of
the field, renaming should help prevent silent merge errors.

Make EXPLAIN better at displaying MRR/BKA:
- Update all .result files
- Extra comments

- "Using MRR" is no longer shown with range access.
- Instead, both range and BKA accesses will show one of the following:
  = "Rowid-ordered scan"
  = "Key-ordered scan"
  = "Key-ordered Rowid-ordered scan"
depending on whether DS-MRR implementation will do scan keys in order, rowids in order,
or both.
- The patch also introduces a way for other storage engines/MRR implementations to
  pass information to EXPLAIN output about the properties of employed MRR scans.

sql/key.cc:
  Fixed compiler warnings about not initialized variables

sql/item.cc:
  If a item_ref is referenced by name, we have to reset it's ref in item_ref::cleanup() as the reference may be to a memory that is freed.
  This happens at least when you have 'HAVING MAX()' in a sub query and you execute 'cond_having->print()' which tries to access items before fix fields.
sql/item.h:
  Added 'reference_trough_name' as a marker for Item_ref that needs to have ref reset during cleanup

Fixed memory leak from HEAP tables that was not deleted properly


BUILD/compile-alpha-ccc:
  Use g++ instead of gcc for linking
BUILD/compile-alpha-debug:
  Use g++ instead of gcc for linking
BUILD/compile-pentium-pgcc:
  Use g++ instead of gcc for linking
BUILD/compile-solaris-sparc:
  Use g++ instead of gcc for linking
BUILD/compile-solaris-sparc-debug:
  Use g++ instead of gcc for linking
BUILD/compile-solaris-sparc-purify:
  Use g++ instead of gcc for linking
sql/item.cc:
  Safety fixes for expr_cache
  Call Item_result:field::cleanup() in Item_cache_wrapper::cleanup()
  More DBUG_PRINT
sql/sql_base.cc:
  Simple optimization for setup_wild
  More DBUG_PRINT
sql/sql_expression_cache.cc:
  Added header
  Removed not needed initialization
sql/sql_lex.cc:
  More DBUG_PRINT
sql/sql_select.cc:
  More DBUG_PRINT
  Fixed memory leak from HEAP tables that was not deleted properly
storage/heap/hp_create.c:
  More DBUG_PRINT

All failing cases were attempts to use connection after
the server was brought down and restarted. Connections 
used client reconnect option.

The reason for failures is the behavior of sockets on Windows:
for a short period after crash (short enough to make the error 
not reproducible under debugger), write to socket on client
side would succeed but subsequent read would fail. 

MYSQL_OPT_RECONNECT does not really help in this case ,
because  in the case given here ,as  mysql_real_query() 
(which can handle reconnect option)  succeeds and 
mysql_read_results() (can't handle reconnect) fails.

The fix is adding --include wait_until_connected_again.inc to 
appropriate places in test. This ensures that read errors are 
caught and connection is recreated.

Fixed problem that fill_record() allocated memory for every call. This could be a problem when doing big unions as memory could be filled up.

sql/sql_base.cc:
  Don't allocate memory in fill_record().
  Fix was to remove the list of tables. This was not necessary as this call is only used with one table.

mysql-test/mysql-test-run.pl:
  Set port-open-timeout to 10 to retry binds
sql/mysqld.cc:
  Write error number for bind failure. (perror can't do that on windows as errno dosn't contain error code)