[GR-75959] Optimize C extension fastcall keyword-name handling.

[graalvmbot]

In macro benchmark c-pydantic-validate, I've observed a lot of time spent in small C API upcalls and a large number of short-lived managed objects that are only created to be handed to native code.

Main changes:

• Convert several hot C API upcalls to static builtins to reduce call overhead.
• Avoid acquiring the GIL for simple allocation upcalls where this is not needed.
• Replace borrowed-value promotion with EnsurePythonObjectNode and avoid boxing primitive values when they can be represented as tagged native handles.
• Create the METH_FASTCALL | METH_KEYWORDS keyword-name tuple directly as a native tuple with interned string elements.
• Implement PyTuple_New in C and allocate a native PyTupleObject instead of creating a managed tuple with native sequence storage. Therefore, adjust tuple checks and type handling to account for native tuples.

Further, this includes small follow-up cleanups around native reference handling and inlining cutoffs that came up while reducing the C API overhead.