Benchmark versioning for comparing perf results

[eellison]

We need versioned benchmark results so perf comparisons are reproducible and diffable across torch branches / dates / hardware:

• A dated/tagged results artifact (e.g. results/<suite>_<hw>_<date>.json) rather than scattered sweep files.
• A compare tool: baseline-vs-branch, per-repro and per-model deltas.
• Pin the torch commit + hardware + corpus tag in the result metadata.
  This is the foundation for the perf-prioritization pipeline (re-bench baseline vs tmp_work, aggregate model-level % improvement).