Train a linear regression model on actual message delivery times to
predict tighter timeouts, replacing worst-case physics estimates.
Features: path length, message bytes, seconds since last RX, flood mode.
Global model with per-contact blending after 10+ observations per contact.
Falls back to existing physics formula when model has insufficient data.

- Clamp ML predictions between physics floor (raw airtime) and ceiling
  (worst-case formula) so model can never produce unsafe timeouts
- Replace hourOfDay feature with secondsSinceLastRx for network activity
- Remove unused _ContactStats.stdDev and dead model persistence code
- Debounce observation writes (2s) instead of writing on every delivery
- Skip recording observations when pathLength is null to avoid corrupting
  training data
- Add comment explaining global (not per-contact) RX time tracking
- Remove notifyListeners from retrain to avoid unnecessary widget rebuilds
- Run dart format