gh-89013: Improve the performance of methodcaller

[eendebakpt]

We improve the performance of methodcaller by using the vector protocol. The performance improvement depends on the use case, but for simple calls the performance doubles. Creation is a bit slower (because the vector call arguments have to be constructed), but the net gain is positive already after a single invocation of the methodcaller

call: Mean +- std dev: [main] 138 ns +- 1 ns -> [patch] 55.5 ns +- 1.6 ns: 2.49x faster
creation+call: Mean +- std dev: [main] 257 ns +- 10 ns -> [patch] 190 ns +- 2 ns: 1.35x faster
call kwarg: Mean +- std dev: [main] 222 ns +- 8 ns -> [patch] 81.4 ns +- 0.4 ns: 2.72x faster
creation+call kwarg: Mean +- std dev: [main] 387 ns +- 11 ns -> [patch] 334 ns +- 5 ns: 1.16x faster

Geometric mean: 1.80x faster

Benchmark script

import pyperf

setup = """
from operator import methodcaller as mc
arr = []
call = mc('sort')
call_kwarg = mc('sort', reverse=True)
"""

runner = pyperf.Runner()
runner.timeit(name="call", stmt="call(arr)", setup=setup)
runner.timeit(name="creation+call", stmt="call = mc('sort'); call(arr)", setup=setup)
runner.timeit(name="call kwarg", stmt="call_kwarg(arr)", setup=setup)
runner.timeit(name="creation+call kwarg", stmt="call = mc('sort', reverse=True); call(arr)", setup=setup)

Note: this is a continuation of #27782. Changes with respect to that PR

• The methodcaller_call was removed in favor of the vector call, but it is still needed for some builds. It is now restored, but less efficient
• Store the full positional arguments as xargs instead of of a slice to avoid creation of an additional tuple
• Additional storage in the methodcallerobject compared to main is now 2 variables PyObject* and a tuple holding the keyword argument names.
• To avoid construction of the vector call arguments during the definition of the methodcaller, one can also defer the construction to the first invocation of the methodcaller. A branch with this approach is main...eendebakpt:cpython:fastmethodcaller_lazy_vectorcall

• Issue: Could operator.methodcaller be optimized using LOAD_METHOD? #89013

[corona10]

close the PR since #107201 is merged.