Fix BLE protocol and MQTT connection bugs

[syscod3]

Summary

• HIGH: Replace blocking time.sleep() with asyncio.sleep() in HADiscoveryPublisher.connect() — the old code stalled the event loop for up to 5 seconds during every MQTT connection attempt. Made connect() and connect_mqtt() async and consolidated into a single asyncio.run() in main().
• HIGH: Fix silent data corruption in _decrypt() — invalid PKCS7 padding (padding_len == 0 or > 16) previously returned garbage bytes; now raises ValueError so callers can handle the error properly.
• MEDIUM: Guard publish_state() and publish_availability() with _connected check — previously called _client.publish() when disconnected, silently losing data.
• MEDIUM: Add is_connected guard in _trigger_update() before write_gatt_char() — prevents an AttributeError/BleakError if the device drops connection between the initial query and the trigger.
• MEDIUM: Replace hard-coded asyncio.sleep(2) after _trigger_update() in read_sensors() with an event-based wait (_response_event.clear() then wait_for(..., timeout=5.0)) — avoids stale signals and gives a proper timeout/log on failure.
• MEDIUM: Validate required sensor keys (mac_address, device_id, local_key) in load_config() — missing keys now cause an early, descriptive error at startup instead of a KeyError during live polling.

Test plan

• Unit-test _decrypt() with padding_len > 16 — expect ValueError
• Run --once against a real SGS01 and confirm MQTT messages appear in HA
• Simulate MQTT disconnect mid-run and confirm publish_state logs error instead of crashing
• Supply a config with a sensor missing local_key and confirm startup exits with a clear error
• Verify ruff passes: python3 -m ruff check pi-fallback/

🤖 Generated with Claude Code