Hot reload / reloadable config

[Niccolum]

Motivation:
The comment in loading/loader.py explicitly defers watchers/polling to the future.
A TTL-based cache exists (loading/cache.py) but does not actively push updates.
Long-running services need to pick up config changes without restarting.

Proposal:
ReloadableConfig wrapper over load():

• get_latest() → T — always returns the most recently loaded instance (immutable,
  no mutation of the previous object, matching hoplite's philosophy).
• subscribe(callback: Callable[[T], None]) — notified after each successful reload.
• Triggers:
  • FixedIntervalTrigger(seconds=N) — background thread polls every N seconds.
  • FileWatchTrigger(paths=...) — OS-level file watch (optional dependency, e.g.
    watchdog) behind an extras group; graceful degradation to polling if missing.
• Reload errors do not replace the current config; they are passed to an optional
  on_error callback.

This is a large feature; consider splitting into two PRs (interval-only first, then
file-watch).

Acceptance criteria:

• Interval trigger reloads on schedule; get_latest() returns new instance.
• Subscriber callback receives the new instance.
• A reload error keeps the old instance; on_error is called.
• File-watch trigger works when watchdog is installed; falls back gracefully.
• Tests for interval + error-handling paths; file-watch tested with mock.
• changes/ fragment.

References: hoplite hoplite-watch, ReloadableConfig, FixedIntervalWatchable,
FileWatcher.