fix: throttle background prefetch and add circuit breaker (#67)

[danmunz]

Fix: Thumbnail prefetch hammering TV into ConnectionFailure spiral

What went wrong with v1.2.0

The background prefetch fix (PR #70) made things worse for users with consistently failing get_thumbnail_list. Nitrowolf's v1.2.0 logs showed:

• POST /api/thumbnails 200 69.59s — response took 69 seconds
• POST /api/thumbnails 200 131.01s — response took 131 seconds
• Dozens of TV op attempt 1/3 failed (ConnectionFailure) per second
• The TV was completely overwhelmed

Root cause: 27 batches of 20 IDs each spawned background prefetch tasks. Each task tried 20 individual _tv_op calls with 3 attempts × 18s timeout each, all serialized through one lock. That's potentially 525 × 3 × 18s = 8+ hours of lock contention. Meanwhile, the frontend's 8 retries also queued up behind the same lock, and each retry re-attempted the failing get_thumbnail_list call (another 8s per try).

Three fixes

1. Circuit breaker on get_thumbnail_list: if it failed in the last 60s, skip it entirely and return the cache state in 0ms (vs 8-60s blocking)

2. Throttled prefetch: single attempt per thumbnail (attempts=1), 0.5s pause between successes, exponential backoff on failure (2s, 4s, 6s…), abort after 5 consecutive failures

3. Single prefetch task: don't spawn additional background tasks while one is already running (prevents 27 prefetch tasks piling up)

Expected behavior for Nitrowolf

• First batch request: get_thumbnail_list fails → returns immediately with fallback: true → spawns ONE prefetch task
• All subsequent batch requests (other 26 batches + retries): circuit breaker skips the TV call → returns cache state in <1ms
• Background prefetch fetches thumbnails one at a time with pauses, gives up after 5 consecutive failures
• No more ConnectionFailure spam, no more 69-131s response times

Closes #67