USB HCI transport teardown — usbi_mutex abort on close

[greateggsgreg]

Closing the USB HCI transport (a Nordic nRF52840 Zephyr dongle, usb:2fe3:000b) aborts

UsbPacketSink.terminate() cancels its OUT transfer but reaps it via a shared semaphore
that a prior completion can pre-release, so a cancellation may go unreaped; and the transfer objects are
freed implicitly by python-libusb1's USBDeviceHandle.close() rather than explicitly after the join.
libusb_free_transfer() on a still-active transfer is illegal and destroys its mutex mid-use

Reproduce

Run repeated BLE sessions over the nRF USB HCI transport and close each; the abort fires at teardown
(exit 134 + the usbi_mutex assertions), including on runs where the BLE work fully succeeded. It is
timing/load dependent in severity but reproduces readily under back-to-back cycles.

[barbibulle]

This cannot be done here, because that's a blocking call.
Waiting for the thread to have finished with join should be done on a worker thread, with asyncio.to_thread(). And in that case, we can get rid of setting event_loop_done prior to returning from run. When to_thread returns, the join is known to have completed.