feat(ci): 4-way matrix slicing with LPT duration-balanced distribution

run_tests_parallel.py:
  - --slice I/N flag (also HERMES_TEST_SLICE env var) runs only the
    I-th slice of N, distributing files across slices by cached
    duration using LPT (Longest Processing Time first) greedy
    algorithm so each slice gets roughly equal wall time
  - Duration cache (test_durations.json): maps relative file paths to
    last-observed subprocess wall time. _save_durations merges with
    existing cache so entries from other slices are preserved.
  - Per-file subprocess timing in progress output + end-of-run
    distribution summary (percentiles, top-10 slowest, <1s/<2s counts)
  - Unknown files default to 2.0s estimate (~P50), spread evenly by LPT

.github/workflows/tests.yml:
  - Matrix strategy: slice [1, 2, 3, 4] with fail-fast: false
  - Each slice restores duration cache from main (stable key, no SHA),
    runs its portion, uploads per-slice durations as artifacts
  - save-durations job (main only, if: always()) downloads all 4
    artifacts, merges into single cache entry for future PRs
  - Timeout reduced from 60min to 30min per slice (~1/4 the work)

Cache design:
  - Stable key (test-durations) not keyed by commit SHA — durations
    are about files, not commits, and SHA-keyed caches miss on every
    new commit and on PR merge commits
  - actions/cache scoping: main's cache is visible to all PRs targeting
    main; feature branches without a cache still work (default 2.0s)
  - No dotfile prefix (upload-artifact v7 skips hidden files)

.gitignore

@@ -18,6 +18,7 @@ __pycache__/web_tools.cpython-310.pyc
 logs/
 data/
 .pytest_cache/
+test_durations.json
 .pytest-cache/
 tmp/
 temp_vision_images/