refactor: consolidate interned Python names

[voltjia]

Summary

• replace seven per-name accessors with one lazily initialized InternedNames instance
• let device and tensor conversion helpers retrieve the shared instance through GetInternedNames()

InternedNames stores 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
• focused CPU+NVIDIA build in an isolated environment with pinned InfiniRT and CUTLASS checkouts
• python -m pytest tests/test_add.py tests/test_gemm.py -m smoke -q -n 1 (12 passed)
• optimized disassembly confirms that both the const-reference implementation and the final internal-getter implementation execute one steady-state singleton guard in tensor conversion; the compiler folds the nested device-helper lookup

Performance

The focused tensor-conversion benchmark used persistent variant workers to exclude interpreter startup from timed regions. Each comparison used 10 independent paired sessions, 15 rounds per session, 50,000 calls per block, alternating B/C/C/B block order, CPU 30 affinity, physical GPU 1 (cuda:0 in the process), and 200,000 session-level paired bootstrap resamples with seed 832. The predeclared non-inferiority margin was +2%.

Comparison	Wall time change	Paired bootstrap 95% CI	Process CPU time change	Paired bootstrap 95% CI
old const-reference implementation (40beaed0) -> final (f05c9e28)	-2.44%	[-5.44%, +0.63%]	-2.46%	[-5.45%, +0.60%]
#832 head (af5d19e2) -> final (f05c9e28)	-1.48%	[-3.77%, +0.59%]	-1.46%	[-3.76%, +0.61%]

All confidence-interval upper bounds are below the +2% regression margin. The point estimates are favorable, but the intervals include zero, so the result is stated as non-inferiority rather than a speedup. The validation-only benchmark binding, harness, and raw results are not part of this PR.