fix(media): run the filesystem index in the background

[ssavutu]

The media index could never complete over HTTP. Fixes the 504 seen on the first
production reindex attempt.

The bug

The index walks ~145k filesystem entries on the real corpus and takes minutes,
but it ran inside the request on r.Context(). The first proxy to give up
killed it:

POST /v1/media/index  status=500  duration_ms=60045

That is Nginx's default proxy_read_timeout of 60s — Cloudflare would have cut
in at ~100s regardless. The walk checked ctx.Err() and aborted cleanly, so the
6,485 rows it had inserted were committed and consistent, but no amount of
retrying could finish the job and the Media → Reindex button was unusable by
construction on a corpus this size.

The fix

• POST /v1/media/index starts the walk on a background context that outlives
  the request and returns 202 immediately.
• GET /v1/media/index reports {running, started_at, finished_at, progress, error}.
• A second POST while a run is in flight returns 409 instead of starting a
  duplicate.
• Runs are capped at two hours, so a wedged filesystem can't leave the job stuck
  running and block every later attempt.
• Progress counts entries walked, not files indexed. The corpus is mostly
  WordPress derivatives that are skipped before any stat, so an indexed-file
  counter appears frozen for long stretches. MediaIndexResponse gains Walked.
• Frontend polls every 2s and shows live counts in the button. A 409 is treated
  as "join the run already in progress" rather than an error.

Unrelated bug this surfaced

refresh() was setSearch(current => current). React bails out of a same-value
setState, so the effect never re-fired and the list silently never reloaded
after an upload or index. It now bumps a token the load effect depends on.

Testing

• go build, go vet, go test ./... clean; tsc -b, eslint, and the
  production build clean.
• New unit tests for the job lifecycle: start/refuse-concurrent/progress/finish,
  failure recording, reset-on-restart, plus 409 and status handlers.
• New integration test (CMS_TEST_DSN-gated, run against real MariaDB 11.8)
  asserting the POST returns 202 in under 2s and the walk completes after the
  request returned, with the derivative skipped and rows actually written by the
  background goroutine.
• Database-level test that the progress callback fires and its final report
  agrees with the returned one, including on trees smaller than the callback
  interval.

Operational note

deploy/README.md updated. Once this is deployed, Media → Reindex works from the
UI; the partial index from the failed run is preserved and the rerun resumes,
skipping what is already there and keeping any alt text.