Fold InputReader into the loop and make UPower construction async

[jwinarske]

Final increment of the event-loop source contract, on top of #22. The BlueZ controllers now run entirely on the single EventLoop thread — no per-device threads and no in-callback blocking D-Bus calls remain.

InputReader is now an EventSource

• The device is opened and its descriptors / feature reports are read in the constructor (non-blocking fd); dispatch() reads and decodes one report per readable event. valid() reports whether initialisation succeeded.
• Its worker thread, stop-eventfd and private epoll are gone. The controller add()s the reader to the loop on device arrival (only when valid()) and retire()s it on removal.

EventLoop::retire()

New primitive: unregisters a source and destroys it at a safe point (after the current dispatch pass). This closes a real use-after-free hazard — a udev "remove" and the reader's own fd (POLLHUP) can be ready in the same poll(), so a handler that reset the reader mid-pass could free an fd still referenced by this iteration's poll set. retire() holds ownership until the next apply_pending().

UPower construction is now async

• UPowerClient fetches properties via GetAllAsync and enumerates devices via callMethodAsync; UPowerDisplayDevice fetches via GetAllAsync. So when a controller with a battery connects, the loop thread is no longer blocked on those round-trips. Pending calls are cancelled by unregisterProxy() if the object is destroyed first.
• UPowerClient keeps its devices_mutex_: it's also used by the standalone src/upower example, which still runs on the connection's own async event-loop thread.

Net −148 lines (the reader threads/eventfds deleted).

Validation

• Runtime (verified): clients start, the loop drives D-Bus callbacks, the async UPower GetAll reply is delivered (onPropertiesChanged: org.freedesktop.UPower logged by the standalone upower_client against live UPower), and SIGTERM shuts down cleanly (exit 0). No regression in the standalone upower shutdown.
• HID input data path: build-verified only — decoding a real report stream needs a physical PS5/Xbox/HoriPad controller, which isn't available in this environment. The decode logic is unchanged from the previous threaded version; only where it runs (loop dispatch() vs a thread) changed.
• clang-format --Werror and clang-tidy-19 clean; full build green (gcc/clang). Two NOLINT(performance-unnecessary-value-param) mark the sdbus-mandated by-value error argument (a const-ref signature fails to compile via function_traits).

Stacked on #22 (now merged); rebased onto main.

Contract complete

With this, the scoped fd() + dispatch() source contract is fully realized for BlueZ: D-Bus, udev, HID input, and signals all run on one EventLoop thread, and every controller mutex introduced back in #18 is gone. The ~14 pure-D-Bus clients still use enterEventLoopAsync() + monitorLoop() — migrating them was called out as optional/low-value in the scoping and is not included.