Improve mirror resilience, selection, and reduce boilerplate

[cmyui]

Summary

• Retry transient errors: _fetch() retries 429/500/502/503 and network exceptions once with 1s backoff
• Rate limit header awareness: reads x-ratelimit-remaining from upstream responses and drains the token bucket when remaining < 10, preventing rate limit trips
• Exponential backoff on circuit breaker: cooldown doubles on each failed probe (30s → 60s → 120s → ... capped at 10min), resets on success — stops broken mirrors (e.g. DNS failures) from wasting hedging slots every 30s
• Failure-aware mirror ordering: adds failure_ema to MirrorHealth and a score() method that penalizes unreliable mirrors (latency * (1 + 3 * failure_rate)), so a fast-but-flaky mirror ranks below a slower reliable one
• Async wait in fallback path: wait_for_availability() briefly sleeps (up to 2s) for rate limiter token refill instead of skipping healthy-but-throttled mirrors
• Deduplicate mirror backends: extracts common _fetch() base method + _extra_headers() hook, reducing ~220 lines of duplicated try/except/404/451 handling across Mino, OsuDirect, and Nerinyan
• Response size limits: rejects responses over 100MB via Content-Length check before buffering
• Remove dead code: get_mirror_weight, BeatmapMirrorScore, MIRROR_INITIAL_WEIGHT were never called

Test plan

• Deploy to production and monitor mirror request logs for correct retry/fallback behavior
• Verify broken Mino mirrors (mino-singapore, mino-germany) back off exponentially instead of probing every 30s
• Confirm score-service beatmap lookup errors decrease

🤖 Generated with Claude Code