[codex] refactor: consolidate interned Python names

[voltjia]

Summary

• replace seven per-name accessors with one function-local InternedNames cache
• pass the cache by const reference through device and tensor conversion helpers
• reuse one cache lookup across vector conversions

The cache retains non-owning py::handle fields, so the intentionally retained interned-name references are not decref'd during static destruction. GetInternedNames() keeps Python C API initialization deferred until a conversion runs with an active interpreter.

Validation

• git diff --check
• clang-format 21.1.8 --dry-run --Werror src/pybind11_utils.h
• fresh focused CPU+NVIDIA builds for both perf: optimize generic tensor conversion #832 head and this commit
• candidate Add smoke: 6 passed
• candidate GEMM smoke: 6 passed
• baseline GEMM smoke: 6 passed

Performance

Baseline: af5d19e254ee14f1a8b7b215354104b56d2877ee (#832 current head)
Candidate: 40beaed0

Both variants were built in fresh virtual environments with the same InfiniRT prefix, CUTLASS checkout, compiler, and focused abs,add,gemm configuration.

Raw cross-process timings on the shared host showed a bimodal CPU P-state (the same conversion path alternated between about 1.53 us and 2.17 us), so those unnormalized timings were not used for the final claim. The causal follow-up used an in-process four-call pybind control with ABBA block order:

• 10 paired A/B cycles
• 15 rounds per process
• 50,000 calls per block
• physical GPU 1, logical cuda:0
• CPU affinity fixed to CPU 30
• 200,000 paired bootstrap resamples, seed 832
• predeclared non-inferiority margin: +2%

Metric	Candidate vs. baseline	Paired bootstrap 95% CI
normalized wall time	-5.65%	[-10.28%, -0.70%]
normalized process CPU time	-5.65%	[-10.27%, -0.71%]

Both confidence intervals are below the +2% regression margin. The validation-only benchmark binding and result files are not part of this commit.