Stop processes whose heartbeats keep failing past the alive threshold

[wintan1418]

Related to #763 (and the same family as #751).

What I found

The core scenario in #763 — a process whose registration gets pruned while it's still alive — already self-heals on main: the next heartbeat raises RecordNotFound, Registrable#heartbeat clears the process and wakes the run loop, shutting_down? returns true via !registered?, the fork exits, and the supervisor replaces it. I've added a regression test proving this end-to-end for the scheduler (start a supervisor with a scheduler, delete the scheduler's solid_queue_processes row while it's running, assert a fresh registration appears).

The remaining hole

That self-heal only fires when the heartbeat gets a clean RecordNotFound. Heartbeats can also fail without confirming anything about the registration — a dropped DB connection after host sleep/resume, SQLite busy errors (the reporter's setup in #763) — and those errors are only reported via the timer task's observer. The process then keeps running unregistered indefinitely: the supervisor never replaces it because it never exits, and from the outside it's a zombie that stops doing visible work. This matches the reported evidence: an alive, supervised scheduler with no solid_queue_processes row and no "Replaced terminated Scheduler" ever logged.

Fix

Track when heartbeats started failing. Once they've been failing for longer than SolidQueue.process_alive_threshold — the exact point at which every other process considers this one dead and prunable — treat it like finding the registration gone: clear the process and wake the run loop so the process stops and the supervisor starts a replacement that can register afresh. Errors are still re-raised so the existing observer reporting is unchanged, and a successful heartbeat resets the tracking, so transient blips don't accumulate.

Testing

• New unit test: a worker whose heartbeats persistently raise ActiveRecord::StatementInvalid shuts down its pool once the failures outlast the alive threshold (fails on main, passes with this change).
• New regression test for the existing pruned-registration replacement path (passes on main and with this change).
• Full worker/scheduler/supervisor/process-recovery test files green on MySQL.