Add per-lane hang timeouts for stuck paper PDF translations

[znsoftm]

A stuck BabelDOC/Google job can occupy the only translate slot for hours (production: 2609.15818 sat ~6h on HTTP 429 with 「166 排队 / 1 翻译中」). This adds a queue watchdog so a hang cannot block the dual-lane pool forever.

Behavior

• While a job is running, wall-clock is measured from persisted started_at.
• Fast lane default 45 minutes; slow lane default 180 minutes (3 hours).
• On timeout: kill the translate_worker / BabelDOC process tree (pidfile + cmdline + descendants), free the in-memory slot immediately, increment persistent timeout_count.
• If timeout_count < max: requeue at the end of the same lane (so other work proceeds) with error timeout (attempt N/3); requeued.
• If timeout_count >= max (default 3): status skipped, error timed out 3 times; permanently skipped. Auto-translate and translate-all will not pick it up again.
• Manual POST /papers/translate for that id resets timeout_count so an admin can retry.
• Applies to in-process runOne jobs and OS workers adopted by reconcileExternal.

Env knobs

Variable	Default	Format
PAPERS_TRANSLATE_FAST_TIMEOUT	45m	Go duration (45m) or integer minutes
PAPERS_TRANSLATE_SLOW_TIMEOUT	3h	same
PAPERS_TRANSLATE_MAX_TIMEOUTS	3	positive integer

Google QPS and default concurrency are unchanged.

Tests

go test ./internal/api/ covers duration/env parsing, requeue + timeout_count, permanent skip, and leaving in-window jobs alone.