gh-139103: Use METH_FASTCALL for tp_new_wrapper

[colesbury]

This addresses one of the bottlenecks when constructing namedtuple instances. With the previous METH_VARARGS, the type argument to __new__ would be included in the arguments tuple, leading to reference count contention.

[colesbury]

A few comments:

• Pulled out from the larger PR gh-139103: Improve namedtuple scaling in free-threaded build #144332, which includes all the fixes needed to address the scaling bottleneck.
• Mostly written by Claude code
• Single threaded performance for the relevant Python __new__ calls is essentially with the noise margin (maybe a tiny bit faster), as measured by https://gist.github.com/colesbury/7c0259e0779aff2cc563c7db298edf42

[vstinner]

LGTM.

I'm not sure what's wrong with the CI.

[colesbury]

I think there are site-wide failures with GitHub Actions hosted-runners: https://www.githubstatus.com/