fix(daemon): survive USB disconnect without crashing and re-attach cleanly

[prodzaha]

Problem

Unplugging the phone while the daemon is running can take the whole process down:

• the USB watcher releases the *gousb.Device handle after the attach callback returns, while relay goroutines may still be inside gousb transfers;
• a forced dev.Close() from a watcher goroutine races with in-flight cgo transfers (gousb._Cfunc_submit), producing SIGSEGV: signal arrived during cgo execution.

Observed twice on macOS 26 (Apple Silicon) + vivo X200 Pro mini: Relay session ended: USB device disconnected is followed by a SIGSEGV dump, launchd restarts the daemon, and the session is stuck until it does.

Changes

• Run all bulk transfers through ReadContext/WriteContext so they observe relay cancellation and never race with a closed handle.
• Track the USB-facing goroutines with a sync.WaitGroup; Relay.Stop waits for them to finish (2s cap) before the device handle is released.
• Drop the watcher goroutine that force-closed the device; context cancellation is enough to unblock the reads.
• Treat LIBUSB_ERROR_NO_DEVICE/NOT_FOUND (and LIBUSB_TRANSFER_NO_DEVICE) as fatal for the session so a disconnect ends the relay instead of looping forever, letting the USB watcher re-attach when the phone comes back.
• Call relay.Stop() from the daemon's session-end path so the handle is only released once no transfers are in flight.

Testing

• go vet ./..., go build ./... clean.
• Runtime (macOS 26.6, Apple Silicon, vivo X200 Pro mini), physical cable unplug/replug:
  • before: SIGSEGV dump, process restart;
  • after: the disconnect is logged, the same process re-attaches, DHCP completes and traffic resumes — no crash line in the log.

Note: this branch alone does not make the vivo X200 Pro mini work end to end — the companion PR fixes the framing/filter side.