bpo-26110: Add CALL_METHOD_KW opcode to speedup method calls with keywords

[Fidget-Spinner]

This is the keyword counterpart for CALL_METHOD (which is only emitted for positional only method calls). The original CALL_METHOD was inspired by PyPy. Not documented there is that PyPy also has a CALL_METHOD_KW for the same purpose.

On microbenchmarks, there is a 1.2x speedup similar to the original patch by Yury for things like X.b(1, 2, c=3, d=4). On pyperformance, it isn't as noticeable -- mostly because pyperformance doesn't use keyword argument calls.

If accepted, docs and What's New (if necessary) will be a later patch.

A side effect of this is that CALL_METHOD should also be slightly faster due to the elimination of a branch.

https://bugs.python.org/issue26110

[Fidget-Spinner]

I'll post pyperformance results here shortly (hopefully by tomorrow).

Edit1: non-scientific timeit microbenchmark results (unfortunately not pyperf timeit :( ) show a roughly 25% time reduction:

 python.exe -m timeit -s "
>> class X:
>>  def foo(self, b, *, c=1, d=2): ...
>> x = X()
>> " "x.foo(2,c=3,d=4)"

# Without the patch (from main):
2000000 loops, best of 5: 146 nsec per loop

# With the CALL_METHOD_KW patch applied:
2000000 loops, best of 5: 107 nsec per loop

Edit2: pyperformance shows no significant change (once again likely because keyword args in method calls aren't common in pyperformance).

[markshannon]

Looks good in general.
A couple of very minor issues

[Fidget-Spinner]

Mark, thanks for taking the time to review this patch. I really appreciate the feedback (it also led to some really interesting discoveries).

I have made the requested changes; please review again

[markshannon]

One unused variable. Otherwise good.

You can change STACK_SHRINK in this PR, or a new one. Up to you.

[markshannon]

Great.

[Fidget-Spinner]

Thanks for all the reviews Mark! I'll watch the buildbots and hopefully nothing bad happens.