Add reproducible performance benchmark suite

[ousamabenyounes]

Summary

• add an easy-to-run benchmark for model-load time, per-run RTF, batch throughput, and peak CUDA memory
• emit reproducible JSON with actual hardware, device, optimization-wrapper, generation, and immutable Hub snapshot metadata
• document benchmark usage, metric definitions, and a verified-results hardware comparison matrix

Validation

• untouched baseline: 56 tests passed
• exact upstream RED: benchmark test collection fails because the runner is absent
• focused GREEN: 27 tests passed with 222/222 benchmark statements covered (100%)
• full GREEN: 83 tests passed
• Black and Flake8 passed for all changed Python files
• source distribution and wheel built successfully; source archive contains the benchmark guide and runner
• two fresh independent Codex review gates passed

Fixes #158