ci: reprocess workflow to redeploy a run's results under an explicit date

[lukekarrys]

Adds a dispatch-only reprocess.yml workflow that re-runs process + deploy from an existing benchmark run's artifacts and writes the output to an explicit date folder, plus a one-line BENCH_DATE override in process-results.sh (defaults to today, so normal runs are unaffected).

Motivation: tonight's full run (31644732609) executed entirely on 2026-08-12 but its process step started at 23:59:45 UTC and date-stamped the output 2026-08-13, leaving the day's regression data in place on the charts. Hand-fixing gh-pages data turned out to be error-prone because each date folder mixes measurement copies with generator-derived files (chart-data.json has four sections, and the history chart reads one of them) — a partial hand-merge left stale derived data twice. This workflow makes the correction reproducible: same artifacts, same generator, CI-authored commit, full paper trail in Actions.

Usage: gh workflow run reprocess.yml -f run-id=<benchmark run id> -f date=<YYYY-MM-DD>. The date must be the calendar day the run's benchmark jobs executed; artifacts are retained 7 days so this only works on recent runs. The deploy uses the same keep_files publish as the daily run, so the target date folder's files are wholesale overwritten with single-provenance output.

Immediate plan after merge: rewrite gh-pages to drop the three hand-fix commits (dc5f3dc, abc9f73, 27ec2d0), then dispatch with run-id=31644732609 date=2026-08-12 so the evening run's complete results replace the morning's regression data with pipeline provenance.