Three interrelated fixes for G522 LIGHTSPEED headset support:
1. Deferred init for silent 0x50 dongles: When the probe fails and
feature discovery returns nothing (dongle silently drops all frames
with device_addr=0x00), return a "pending" CenturionReceiver instead
of falling through to the broken create_device direct-device path.
The listener thread starts reading; when the first unsolicited frame
arrives, _unwrap_centurion_frame learns device_addr, the notification
handler detects the pending state, and re-runs feature discovery with
the correct address — finding the bridge and creating the child device.
2. Centurion protocol version floor: The G522 dongle reports protocol
1.1 (major=1, minor=1), which routes all protocol < 2.0 gates into
HID++ 1.0 code paths (battery register reads, etc.) that crash with
INVALID_SUB_ID_COMMAND. Centurion devices always use HID++ 2.0
features, so the protocol property now returns 2.0 as a floor for
any device with centurion=True.
3. CI segfault fix: Mock probe_centurion_device_addr in the two
Centurion device tests so fake handles never reach the real hidapi
C extension on macOS.
e91cbff361