Add full end-to-end Ampere validation runner

[groxaxo]

Summary

Adds scripts/validate_gptq_pro_ampere.sh, a single end-to-end validation command for physical Ampere systems.

The runner:

• automatically selects the least-busy GPU or accepts an explicit physical GPU index;
• refuses a busy GPU unless explicitly overridden;
• records Git, Python, CUDA, driver, package, and GPU metadata;
• runs isolated qweight/dispatch/source contract tests;
• compiles the standalone validator for sm_80/sm_86/sm_87 plus PTX, or native-only by request;
• executes the numerical CUDA validator;
• runs compute-sanitizer memcheck and optional racecheck;
• builds/imports the real versioned PyTorch CUDA extension;
• benchmarks AUTO, GEMV, Ampere, and legacy paths across decode/prefill shapes;
• enforces configurable numerical gates;
• reports AUTO-vs-legacy timing without falsely requiring a speedup by default;
• writes validation-report.json, SHA-256 artifact hashes, and VALIDATION_PASSED only after all enabled correctness checks pass.

Usage

bash scripts/validate_gptq_pro_ampere.sh
bash scripts/validate_gptq_pro_ampere.sh --quick
bash scripts/validate_gptq_pro_ampere.sh --gpu 2 --racecheck

The permanent kernel CI is extended to run bash -n and the runner's --help path.