Coalesce reload broadcasts within a quiet window

[joeljunstrom]

Listen + macOS fsevents commonly fires several writes for a single logical change: esbuild rewriting a bundle plus its sourcemap, editors performing atomic-rename saves, asset pipelines touching siblings. Each write currently flows through Hotwire::Spark::Change#broadcast and reaches the browser as a separate reload frame, producing visible flicker and redundant work.

This behaviour / issue becomes apperant when running Falcon in development due to the increased concurrency.

Routes broadcasts through Hotwire::Spark::Debouncer, a trailing-edge debouncer keyed on [action, canonical_changed_path]. Each event extends a quiet-window deadline; once no event has arrived for debounce_window seconds (default 0.1), pending entries flush as one broadcast per key. The shared state is mutex-guarded; the actual broadcast runs outside the lock so ActionCable latency does not serialize incoming watcher events. Per-block rescue keeps one failing broadcast from dropping siblings in the same flush.

This is part of an ongoing effort to make rails development be as nice as usual when using Falcon / fiber concurrency in local development.

• Fix ShareLock ownership under :fiber isolation rails/rails#57423
• Release reloader share on rack hijack in ActionDispatch::Executor rails/rails#57425
• Ensure socket is closed even when transmit raises in Connection#close anycable/actioncable-next#17

Note that #109 is required in order for Spark to work under falcon