refactor: minimize Python C API usage in tensor conversion

[voltjia]

Summary

• represent interned attribute names as pybind11::handle instead of exposing raw PyObject * values throughout the conversion helpers
• replace the custom PyObject_GetAttr wrapper with public pybind11::getattr
• keep the Python C API fast paths that do not have a zero-overhead public pybind11 equivalent

Motivation

PR #832 gets most of its speedup from cached attribute names, fused no-argument method calls, and fast metadata conversion. This follow-up narrows direct Python C API usage without giving those gains back.

A broader prototype also replaced PySequence_Fast_GET_* and direct UTF-8 access with public pybind11 iteration/casters. It regressed the 16x16 GEMM B -> C comparison by 4.25% (95% CI: +1.43% to +6.56%), so this revision deliberately retains those measured hot paths. The remaining direct calls are limited to name interning, fused no-argument calls, integer extraction, fast sequence access, UTF-8 views, and their error handling.

Validation

Environment:

• NVIDIA A100-SXM4-80GB, physical GPU 3 exposed as cuda:0
• process pinned to CPU 24, with a 1 second CPU ramp before each timed case
• Python 3.12.3, pybind11 2.13.6, Torch 2.10.0a0+nv25.12
• InfiniRT 31b2a32c3e0b8eb2ef9da634c3ab934ce87e6af3
• 5,000 warmup calls, 20,000 calls per round, 9 rounds per process
• 6 balanced A/B/C process-order cycles; process median is the statistical unit
• 200,000 paired bootstrap resamples for 95% confidence intervals

A is PR #832 base 2cfb526, B is PR #832 head 50f979df, and C is this branch.

Case	A -> B	B -> C	A -> C
tensor conversion, 16x16 FP32	-35.94% [-42.00%, -31.57%]	-0.42% [-2.77%, +1.98%]	-36.21% [-42.23%, -31.92%]
Add, 13x4 FP32	-24.14% [-31.24%, -19.14%]	+0.57% [-1.16%, +1.97%]	-23.70% [-30.25%, -19.80%]
GEMM, 16x16 FP32	-18.69% [-26.30%, -13.99%]	-0.77% [-2.21%, +0.82%]	-19.31% [-27.31%, -14.00%]

The upper bound of the B -> C regression interval stays below the preselected +2% non-inferiority margin for all three cases.

Checks:

• focused CPU/NVIDIA wheel build with abs,add,gemm
• pytest tests/test_gemm.py -m smoke -q -n 1: 6 passed
• ruff check tests/conftest.py
• clang-format 21.1.8 --dry-run --Werror src/pybind11_utils.h
• git diff --check