The BLE provider never listened for the Web Bluetooth
`gattserverdisconnected` event, so when the Spark dropped its BLE
link the app would keep `isConnected = true`, the UI would still
report "connected", and every subsequent command silently failed.

This change wires up a disconnect-detection + single-attempt
auto-reconnect path while keeping the existing manual connect/
disconnect flow untouched.

Behavior:
  * BleProvider attaches a `gattserverdisconnected` listener after
    services are wired up. On unexpected drop it tears down local
    state, detaches the notification listener, clears the receive/
    send queues, fails any pending ack waiters, and invokes
    `onDisconnected`.
  * A new `intentionalDisconnect` flag plus listener removal in
    `disconnect()` keeps user-initiated disconnects from triggering
    the auto-reconnect path.
  * `BleProvider.reconnect()` re-uses the cached `BluetoothDevice`
    handle and re-runs service discovery, so the user is not
    re-prompted with the device picker.
  * `SparkDeviceManager` exposes `onConnectionLost` and `reconnect()`
    and now tracks its receive-loop `setInterval` id so reconnect
    cycles don't stack additional 50ms intervals.
  * `DeviceContext` reacts to connection loss by emitting a
    'disconnected' event to the UI and kicking off a single
    `attemptReconnect()` after a 1.5s settle delay. A monotonic
    `reconnectGeneration` token cancels stale attempts when the
    user takes a competing action (manual scan/connect) so the
    UI state is not clobbered by a superseded attempt.
  * `DeviceViewModel` handles the new 'disconnected' /
    'reconnecting' events alongside the existing 'connected' /
    'failed' ones; the status dot now reflects reality.

Correctness fixes uncovered during review:
  * `connect()` previously left `isConnected = true` if service
    discovery failed after the GATT link was up. A later reconnect
    would short-circuit and report success against an unusable
    connection. Now rolled back fully on failure.
  * `beginQueuedReceive()` registered an anonymous
    `characteristicvaluechanged` listener with no removal path,
    leaking across reconnects. The bound listener and its target
    characteristic are now tracked and removed on every disconnect.
  * `write()` set `isSendQueueProcessing = true` without a
    `try/finally`, so a write that threw (e.g. mid-drop) would
    leave the queue permanently "processing" and stall all future
    commands. Now wrapped in try/finally.
  * `SparkDeviceManager.startReceiver()` called `setInterval` with
    no reference, so every reconnect would add another loop on top
    of the existing one. The interval handle is now stored and
    cleared on disconnect/reconnect.

Out of scope (intentional follow-ups):
  * Multi-attempt backoff (currently exactly one auto-attempt).
  * Aborting an in-flight Spark 2 chunked preset upload when the
    underlying BLE link drops mid-upload.
  * Heartbeat-based liveness for stalls that don't fire
    `gattserverdisconnected` (some Windows BLE stacks).

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>