fix(lora): make async init recoverable and thread-safe

[Forairaaaaa]

The async-init rework used a detached thread plus a one-shot, no-retry path. A transient cold-boot / EXT5V failure would wedge the page as "LoRa unavailable" with no recovery, and re-entering the page could run two init threads against the global backend state at once.

• Keep the init thread joinable and join it in the destructor and before each new attempt, so the global hardware state is never touched by two init threads concurrently.
• Retry failed initialization with a 3s backoff instead of giving up.
• Split the misleading "LoRa is still initializing" send status into a distinct "LoRa unavailable" message.