Run the BlueZ controllers on a single EventLoop, drop the mutexes

[jwinarske]

Step 2 of the event-loop source contract, building on the EventLoop from #21. Moves the three BlueZ game-controller clients onto a single-threaded event loop and deletes the concurrency that PR #18 had to harden.

What changed

• UdevMonitor is now an EventSource — the netlink monitor is set up in the constructor; fd() exposes the socket and dispatch() drains queued device events. Its worker thread and self-pipe are gone.
• SignalSource (new, in src/utils/) — a signalfd EventSource that stops the loop on SIGINT/SIGTERM. It's constructed before any thread starts so the signals are blocked process-wide and delivered only through the loop.
• The three controller mains replace enterEventLoopAsync() + monitorLoop() with EventLoop::run(), registering the controller (itself a UdevMonitor) and a SignalSource.
• All the mutexes are removed from the three controllers (adapters_/devices_/input1_/upower_clients_ and input_reader_). The D-Bus object-manager callbacks and the udev callback now run on the same loop thread, so the state they share needs no locking — this retires the exact cross-thread surface behind PR Fix reliability, security, and performance issues across clients #18's race fixes. Net −122 lines.

Validation

• Runtime (verified): each client starts, EventLoop::run() drives the D-Bus callbacks on the loop thread (object-manager/property callbacks observed firing), and SIGTERM is delivered through the signalfd — SignalSource: received signal 15, stopping → clean exit 0. Caught and fixed a real bug along the way: with the old main ordering, spdlog's periodic-flush thread was started before signals were blocked and took the default SIGTERM disposition, killing the process; constructing SignalSource first fixes it.
• udev source: build-verified; dispatch is the identical EventSource mechanism the SignalSource exercises at runtime (couldn't inject a real uevent in the CI/sandbox — udevadm trigger is unavailable there).
• HID input data path: build-verified only — exercising it needs a real PS5/Xbox/HoriPad controller. InputReader is unchanged here (still on its own read thread), so that path is as it was in Fix reliability, security, and performance issues across clients #18.
• clang-format --Werror and clang-tidy-19 clean; full default build green.

Scope / next

This increment does the structural unification (udev + D-Bus + signals on one loop). It intentionally does not yet:

• fold InputReader into the loop as a source (it keeps its read thread; it only logs and touches no shared owner state), or
• convert the in-callback UPower construction (GetAll + EnumerateDevices) to the async sdbus API.

Those two are the next increment — the async conversion is the approach chosen for keeping the single loop from blocking on in-callback D-Bus round-trips. Until then, UPower client construction runs synchronously on the loop thread (a brief stall on controller connect), which is acceptable for these monitor daemons.

Stacked originally on #21 (now merged); rebased onto main.