perf: optimize generic tensor conversion

[baominghelly]

Summary

• Cache/intern Python tensor metadata attribute names used by TensorFromPybind11Handle.
• Use CPython no-argument method, sequence, integer, and UTF-8 APIs to reduce temporary pybind11 objects and string/vector copies in the generic duck-typed tensor conversion path.

Motivation

TensorFromPybind11Handle runs for every tensor argument passed through the generated eager Python bindings. The previous implementation repeatedly created attribute-name objects and converted Python strings/sequences through generic pybind11 casters.

On an NVIDIA A100-SXM4-80GB, an alternating A/B run of the exact patch used FP32 16x16 GEMM with implementation_index=0, 200 warmups, 2,000 calls per repetition, 15 repetitions per process, and four runs per variant:

Metric	Baseline	Candidate	Improvement
Torch tensor conversion	4.279 us	2.586 us	1.693 us (39.6%)
Generic duck tensor conversion	3.693 us	2.546 us	1.147 us (31.1%)
GEMM end-to-end latency	19.889 us	15.891 us	3.997 us (20.1%)

Type of Change

• feat — new feature / new operator / new platform
• fix — bug fix
• perf — performance improvement (no behavioral change)
• refactor — code restructuring without behavior change
• test — adding or fixing tests only
• docs — documentation only
• build / ci — build system or CI configuration
• chore — tooling, formatting, or other non-code changes
• Breaking change (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

CUDACXX=/usr/local/cuda/bin/nvcc CMAKE_BUILD_PARALLEL_LEVEL=32 \
SKBUILD_BUILD_DIR=build-pr-smoke-cuda13 \
python -m pip install '.[dev]' --no-build-isolation --no-deps \
  --config-settings=cmake.define.INFINI_RT_ROOT=/data/shared/baoming/workplace/new_ops/infinirt-prefix \
  --config-settings=cmake.define.INFINI_OPS_SMOKE_BUILD=ON \
  --config-settings=cmake.define.WITH_NVIDIA=ON \
  --config-settings=cmake.define.WITH_CPU=OFF

Successfully built InfiniOps
Successfully installed InfiniOps-0.1.0

CUDA_VISIBLE_DEVICES=4 python -m pytest tests -m smoke -q
2 failed, 91 passed, 19 skipped, 13008 deselected in 16.35s

The two failures are the existing FP32 GEMM tolerance cases for implementations 0 and 1. Both failures reproduce with the same output values on an independently built, unmodified master at 26228d8; the candidate and baseline GEMM smoke subsets each report 2 failed, 8 passed.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Build passed; 91 passed, 2 baseline-reproduced failures, 19 skipped	Full suite not run
Iluvatar	Yes	Not run - no hardware available	Shared generic conversion path
MetaX	Yes	Not run - no hardware available	Shared generic conversion path
Cambricon	Yes	Not run - no hardware available	Shared generic conversion path
Moore	Yes	Not run - no hardware available	Shared generic conversion path
Ascend	Yes	Not run - no hardware available	Shared generic conversion path

Full `pytest` output (optional)

Not run - this is a draft performance PR; NVIDIA smoke validation is recorded above.

Benchmark / Performance Impact

• Harness: out-of-tree tensor-conversion profiler, intentionally excluded from this PR.
• Hardware: NVIDIA A100-SXM4-80GB; no foreign GPU process was observed during the alternating A/B run.
• Workload: FP32 16x16 GEMM, explicit implementation index 0.
• Method: 200 warmups; 2,000 end-to-end iterations; 20,000 conversion iterations; 15 repetitions; four alternating runs per variant; median across run medians.
• Result: generic tensor conversion improved by 31.1%, direct Torch tensor conversion by 39.6%, and GEMM end-to-end latency by 20.1% (3.997 us).

Notes for Reviewers

• The optimization remains framework-neutral and preserves the existing duck-typed tensor contract (data_ptr, shape, dtype, device, and stride).
• Interned Python attribute names intentionally retain one process-lifetime reference to avoid destruction after Python interpreter finalization.
• This PR does not include ATen tensor reuse, THPVariable_Unpack, Torch wrapper generation changes, CMake toggles, profiling hooks, benchmark scripts, or profiling logs.
• Non-NVIDIA backend validation and the two pre-existing GEMM tolerance failures remain follow-up items before marking the PR ready.