feat: persist per-attempt job runs

[giscebot]

Summary

• persist a job_runs record for every claimed attempt and close it on done, blocked, requeued, or cancelled transitions
• make completed runs the source of truth for runtime totals and percentiles while preserving the existing job summary fields
• assign each run to the local day with the largest elapsed overlap, using the start day for ties and deriving the month from that day
• expose run history in job details and update the dashboard counter to report runs

Migration and retention

• schema initialization backfills at most the last recoverable legacy interval as historical with is_estimated=1
• historical attempts that cannot be reconstructed are not invented
• runs follow their parent job lifecycle through ON DELETE CASCADE; they are not independently pruned so historical usage does not change silently

Tests

• .venv/bin/pytest -q — 392 passed
• npm test — 60 passed
• npm run build

Closes #227

Requested by: @ecarreras

[pilipilisbot]

Approved. I reviewed the queue/run lifecycle changes against #227 and did not find blocking issues.

What I checked:

• claim_next() creates a per-attempt job_runs row and keeps jobs as the compatibility summary.
• finish(), requeue_running(), block_running(), mark_cancelled(), and unlock_stale() close the active run with the expected result.
• Legacy schema initialization backfills only the recoverable interval and marks it estimated.
• Runtime metrics now use completed job_runs, preserve work/review grouping, and assign cross-midnight runs by largest local-day overlap with start-day tie behavior.
• The dashboard detail exposes run history, and packaged resource coverage includes the new table.

Validation:

• Local focused suite: /tmp/gab-pr236-venv/bin/pytest -q tests/test_queue.py tests/test_backend.py tests/test_packaged_resources.py -> 111 passed, 1 warning.
• Local full Python suite: /tmp/gab-pr236-venv/bin/pytest -q -> 392 passed, 1 warning.
• GitHub checks are green for dashboard, pytest 3.11, and pytest 3.12.

I did not rerun the dashboard npm tests locally because this review worktree has no node_modules; the PR's dashboard check is green.