SmoothQuant reference implementation (PyTorch W8A8)

IMPORTANT methodology point for every SmoothQuant quality delta in this repo: the accuracy numbers come from this PyTorch harness (evaluated with lm-eval-harness), while the 1.51x/1.56x speedups come from the FasterTransformer INT8 integration. Accuracy and speed in that paper are NOT from the same runtime. O1-O3 are different schemes, not different strengths of one scheme: O3 is per-tensor static and is the one that loses accuracy on OPT-175B. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2024-07-12 (26.7 months before this record), NO GitHub release or tag found, 4 contributors (distinct committers including anonymous), open_issues_count 72 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), not archived, 1692 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set. ||| DEPLOYMENT: NO EVIDENCE OF PRODUCTION DEPLOYMENT FOUND as of 2026-10-04 - repositories, READMEs, docs and issue trackers were searched. See [[ev-production-evidence-is-scarce-and-nearly-always-single-site]] for the grading scheme; note that for an artifact-evaluation or paper-code release, non-deployment is documented intent rather than a negative finding.