feat: reactive watch API (watch* subscriptions)

[edisdev]

Summary

Adds a reactive watch API so consumers can subscribe to a kind of device
information and get a callback whenever the value changes, instead of polling a
getter on a timer. The current value is delivered immediately, then on every
change.

import { watchBattery } from "tauri-plugin-device-info-api";

const unwatch = await watchBattery((b) =>
  console.log(`Battery: ${b.level}% (charging: ${b.isCharging})`)
);
// await unwatch(); // stop when done

New functions: watchBattery, watchNetwork, watchStorage, watchDisplay,
watchDevice — each returns a Promise<UnwatchFn>. New exported types:
WatchOptions (intervalMs) and UnwatchFn.

How it works

The public API and event names are identical across platforms; the engine per
kind differs:

• Native, event-driven on macOS — battery (IOKit), display (Core Graphics),
  network (SystemConfiguration). CPU stays idle between changes; updates arrive
  the instant the OS reports them.
• Change-detecting polling fallback for kinds/platforms with no OS event
  (storage and device everywhere; all kinds off macOS). Emits only on a real
  change.

Monitors are reference-counted: one per kind, started on the first
subscriber and torn down on the last.

Performance

• Per-kind polling defaults/floors so expensive getters aren't hammered:
  device 60s (min 10s), storage 10s (min 1s), others 2s (min 250ms).
  device shells out to system_profiler on macOS and rarely changes, so it's
  no longer read every couple seconds.
• Poller sleeps on a condvar for the full interval and wakes immediately on stop
  (no 100ms busy-wake), so the CPU can idle.
• Change events emitted by reference — no deep clone per snapshot.

Correctness / safety

These came out of a self-review of the new backend code:

• Race-free macOS run-loop teardown — stop is signalled through the run
  loop via a CFRunLoopSource, so a stop that races startup can't be lost
  (previously CFRunLoopStop before CFRunLoopRun could hang stop_watching).
• stop_watching no longer holds the shared watcher lock while joining the
  monitor thread, so a slow teardown can't block other kinds' subscribe/stop.
• OS callbacks contain panics so they can't unwind across the C ABI and
  abort the process.
• The JS watch() removes its event listener if the backend fails to start; the
  example surfaces a failed start instead of leaving an unhandled rejection.

Testing

• cargo test — added watcher lifecycle tests (spawn → emit → stop → join → free)
  for every kind, including the native macOS paths; ran the start/stop test in a
  loop to confirm no teardown hang.
• cargo clippy clean.

Docs & versioning

• New Reactive Watch API reference page + guide/examples/README updates.
• Version bumped 1.0.1 → 1.1.0 (backwards-compatible), new CHANGELOG.md.

Commits

Split into reviewable layers: backend watcher · JS bindings · example app ·
docs · release/changelog.

🤖 Generated with Claude Code