GeezerVison

Changes look good to me. On a local build the new code did not have a tail call, while current main did (MSVC compiler). I used Claude to let the argument clinic generate slightly easier code to compile. A godbolt link that compares main, this PR and the branch linked above: https://godbolt.org/z/5WodhW3nb

These changes are minor though (and compiler dependent, one could argue MSCV just needs to do a better job), so I am fine with merging this as is.

Appreciate the numbers. Lines up with some of what I've been finding: A lot of Argument Clinic binding performance is in compiler inlining and optimization.

I've been working on getting str, list, and int to be able to use AC Vectorcall and they are all slower because the hand written manually inlines important cases. Compilers see "two callers" + "too much complexity" and don't do the inlining most the time. AC doesn't know there is something it could inline (ex. here's how to get the global constant empty bytes for zero-arg, here's the one-arg from type A, etc.). Have some ideas on how to help AC have that information and be able to provide more type-specific overloads for the JIT as well potentially... but very prototype.

Hoping as keep iterating on AC will get some large systemic improvements. AC Vectorcall is faster when there are keyword args but often slower for positional arguments in conversions currently. When there wasn't a Vectorcall before things speedup especially if can unlock additional optimizations. Speedup depends on how much overhead the arg conversion was.