instead of re-calculating it for every BB

profiling ftw

Don't use ValueMap::operator[], since that is much more
expensive than lookup().

I think `make check_dbg` is clean again.

Most of these need to return borrowed references in order to match the C API.

If we find a type that
- we think we can reason about statically
- does not have an __iter__ but can be iterated via __getitem__

we previously just bailed saying that we know it doesn't have an __iter__ method
(instead of calling getiterHelper like we should).  This bug exists on master as well,
but we just don't run into it that often (especially since it's llvm-tier-only).

With some prefetching.

Now that they are only a small number of sizes (powers of two),
can use a freelist pretty easily.  Especially because most of the
allocations will come from places where we can precompute the freelist
index.

About a 5% improvement on raytrace.py

Our old allocator supported very fast realloc() calls (amortized constant
time), and we relied on that by doing "realloc(array, old_size + 1)" every
time we wanted to grow the array.

Instead, now do "if (power_of_two(old_size)) realloc(array, old_size * 2)"
Somehow this is 15% faster overall.

Refcounting added a second way to represent "no attrs" -- NULL as a hcls.
This was causing guards to fail since a check would be written for one
but then it would fail when objects came in with the other.

So canonicalize to using NULL as the "no attributes" hcls.

It can still be helpful to turn them off, but now that the refcounter non-determinism
has been fixed I think it's time to have these back on.

this let's minibenchmark/raytrace.py run with enabled interpreter + OSR (but disabled bjit)

(we had already copied their int freelist)

I guess it was originally turned on for the extra debugging it provides,
but it also a decent amount faster than system malloc/free
(though I think we could do better still).

Not sure how I missed the dict.cpp file

Put a "expected: reffail" on any test that the tester thought was failing
(which includes a few that currently fail due to the llvm-only mode)

Used a different word (vs "fail") so that we can see whether it was failing
before refcounting or not.