feat: durable cache so finished sessions survive restarts

[JustMaris]

Summary

OpenF1's free tier returns 401 Live F1 session in progress for everything (past sessions included) for a stretch around every live session - observed starting ~15 min before a session and lifting ~30 min after. Finished sessions were already kept in memory so a lockout couldn't blank them, but a restart wiped that: a deploy landing inside a lockout left the Tyre Usage and Latest Session tiles empty until it lifted.

• cache.getDurable / setDurable keep never-changing data in memory and, when CACHE_DIR is set, mirror it to disk as small JSON files (atomic temp-file + rename; ignored after 30 days). tyre_usage and latest_session use it for finished sessions; anything else immutable can too.
• Image: /data owned by the non-root user (65532) and CACHE_DIR=/data by default. Mount a volume there to persist; without one it still works, it just doesn't outlive the container.
• Startup probe: logs a warning if the directory isn't writable, instead of silently not persisting.
• README + docker-compose example updated.

Test plan

• gofmt / go vet / go test ./... clean; new test: a "restarted" instance (empty memory, same dir, upstream locked out) still serves tyre and latest-session data, while one without a dir shows upstream_error
• Built the real scratch image and ran it: fresh named volume -> starts clean; no volume -> starts clean; read-only mount -> logs the "CACHE_DIR is not writable" warning
• Not exercised against a real OpenF1 lockout end-to-end; check after deploy that a restart during a lockout keeps the tiles populated

Cluster side (PVC + mount): see the companion PR in drumandbytes/cluster.