Type the gateway's zigpy device handling against BaseDevice instead of blanket mypy suppressions #931

Follow-up from the review thread on #914 (zigpy 2.2.0 → 2.3.0).

Since zigpy 2.3.0 (zigpy/zigpy#1874 split BaseDevice into ZigbeeDevice and GreenPowerDevice), ControllerApplication.devices and get_device() return BaseDevice, and the device resolver type is Callable[[BaseDevice], BaseDevice]. The ZHA gateway still annotates these as ZigbeeDevice. mypy only stays green because zha.application.gateway has blanket arg-type and return-value suppressions in pyproject.toml.

Running mypy without the per-module overrides (the run tools.regenerate_mypy_ignores does internally) on dev at 0ea5e1a shows three errors from this:

• gateway.py:252: device_resolver=DEVICE_REGISTRY.resolve passes a Callable[[ZigbeeDevice], ZigbeeDevice] where Callable[[BaseDevice], BaseDevice] | None is expected.
• gateway.py:302: _find_coordinator_device returns the BaseDevice from get_device() as a ZigbeeDevice.
• gateway.py:309: load_devices passes a BaseDevice from application_controller.devices to get_or_create_device, which takes a ZigbeeDevice.

The same two codes also cover four unrelated gateway errors (lines 217, 285, 409 and 423), so the suppressions can't just be dropped. The count comments are stale too: arg-type is 5 (comment says 3) and return-value is 2 (comment says 1).

There is no runtime impact today. zigpy 2.3.0 defines GreenPowerDevice but never instantiates it: commissioning isn't implemented, and GP frames from unknown devices are logged and dropped. Every device the gateway sees is still a ZigbeeDevice.

This is an issue rather than a PR because the right fix depends on how ZHA ends up representing GP devices, and that work is in flight:

• Base ZGP device support #868 (draft, base ZGP device support) changes DEVICE_REGISTRY.resolve and Gateway.get_or_create_device to take a BaseDevice. At 60919c2, with the overrides removed, that clears the resolver and load_devices errors. The coordinator one stays (line 308 there), and four new ones show up that the gateway override hides: line 697 (get_device returns a BaseDevice | None from _devices as Device | None), lines 775 and 782 in async_device_initialized, where the BaseDevice returned by get_or_create_device goes into _async_device_rejoined / _async_device_joined, and line 266, where that BaseDevice is assigned to coordinator_zha_device: Device | None (an assignment error). _async_device_rejoined and _async_device_joined take a ZHA Device and call async_configure(), which only the Zigbee device class has.
• zgp: add the Green Power manager, device, and proxy runtime zigpy#1848 (open, GP manager runtime) keeps commissioned GPDs in its own green_power registry with its own events, not in app.devices or the device_initialized listener event. So it isn't settled yet which gateway paths a GPD will reach.

Adding isinstance(..., ZigbeeDevice) skip guards on dev now would contradict #868, which creates ZHA devices for GPDs instead of skipping them.

Tasks:

• _find_coordinator_device: narrow with assert isinstance(zigpy_coordinator, zigpy.device.ZigbeeDevice). The coordinator is always a Zigbee device (zigpy's get_device(nwk=...) only matches ZigbeeDevices), zha/application/platforms/virtual.py already asserts the same (line 211), and this doesn't depend on the GP design.
• Once the GP device model is settled (Base ZGP device support #868 and the zigpy side): make the gateway paths that receive zigpy devices (the resolver, load_devices, the device_* listener callbacks) either accept BaseDevice or narrow explicitly. A GPD that reaches a Zigbee-only path (configure, re-interview, groups) should then be a mypy error, not a suppressed one.
• Regenerate the overrides with python -m tools.regenerate_mypy_ignores afterwards. On dev today the regen also updates zha.application.platforms.sensor (arg-type 12 → 11, attr-defined 5 → 2, operator 4 → 2), which moves it and the gateway entry in the count-sorted list.
• Optional: fix the four unrelated errors as well, so arg-type and return-value can be removed from the gateway override entirely.

# dev at 0ea5e1ad, zigpy 2.3.0 (uv.lock), worktree venv
.venv/bin/python -m mypy zha/                      # Success: no issues found in 63 source files
# unsuppressed run via tools.regenerate_mypy_ignores.run_mypy(), gateway arg-type/return-value only:
gateway.py:217 [return-value]  tuple[type[ControllerApplication], ...] vs tuple[ControllerApplication, ...]
gateway.py:252 [arg-type]      device_resolver Callable[[ZigbeeDevice], ZigbeeDevice] vs Callable[[BaseDevice], BaseDevice] | None
gateway.py:285 [arg-type]      request_priority: PacketPriority | None vs int
gateway.py:302 [return-value]  BaseDevice vs ZigbeeDevice
gateway.py:309 [arg-type]      get_or_create_device: BaseDevice vs ZigbeeDevice
gateway.py:409 [arg-type]      DeviceJoinedDeviceInfo ieee: EUI64 vs str
gateway.py:423 [arg-type]      RawDeviceInitializedDeviceInfo ieee: EUI64 vs str
# same run on #868 at 60919c2 (contains #914, zigpy 2.3.0): 252 and 309 gone; 308, 697, 775, 782 present, plus 266 [assignment]