Fix duplicate overload type resolution

[ZzEeKkAa]

Fix duplicate overload type resolution

Summary

@overload / @overload_method implementation functions are re-executed once per
distinct compiler-Flags context in which a call is resolved. Because the overload
body is where user code does its (sometimes very expensive) work, this doubles that
work for every overload that is resolved both during an enclosing kernel's type
inference and during a device-function compile.

This MR memoizes the overload-function result on the typing context + argument
types only — the inputs the overload actually depends on — so the body runs at
most once per distinct signature, while leaving the existing flag-sensitive
_impl_cache untouched.

Root cause

_OverloadFunctionTemplate._get_impl keys its implementation cache on the active
flags:

flags = targetconfig.ConfigStack.top_or_none()
cache_key = self.context, tuple(args), tuple(kws.items()), flags

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[ZzEeKkAa]

/ok to test

[copy-pr-bot]

Auto-sync is disabled for ready for review pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.