Add single-threaded EventLoop that multiplexes D-Bus and fd sources

[jwinarske]

First step of the event-loop source contract that was scoped earlier. This PR adds the machinery only — no client is converted yet.

What

• EventSource — fd() + events() + dispatch(revents), invoked on the loop thread.
• EventLoop — drives an sdbus::IConnection via its external-loop poll data (getEventLoopPollData() / processPendingEvent()) alongside a set of EventSources, all on the thread that calls run(). This is the alternative to enterEventLoopAsync(): because D-Bus callbacks and device/fd I/O run on one thread, state shared between them needs no locking.
  • stop(code) is async-signal-safe and thread-safe (writes an eventfd to break poll()), so it can be driven from a signal handler or another thread.
  • add() / remove() defer to the next iteration, so they are safe to call from within a dispatch() (controller plug/unplug does exactly this). Removed sources must outlive the iteration — documented on the method.

Why

It's the foundation for retiring the per-client worker-thread + std::mutex pattern in the BlueZ clients — the same fragmentation that PR #18's races came from. sdbus-c++ supports exactly this via getEventLoopPollData()/processPendingEvent(), so the D-Bus connection, UdevMonitor, and each InputReader can eventually share one loop.

Validation

Opt-in self-test (-DSDBUS_CPP_EXAMPLES_BUILD_TESTS=ON, wired into CTest) drives the live system bus with two fd sources on a single thread and asserts:

1. a semaphore-eventfd source is dispatched once per token (3/3),
2. an async ListNames reply is delivered by run() — proving the loop drives processPendingEvent (125 names, ok),
3. stop() unblocks run() and returns 0,

with a timerfd as a hard deadline so it can never hang. Passes repeatably; ctest 100%. Verified clean under clang-format --Werror and clang-tidy-19.

Notes

• The test is opt-in and not built by CI (the option is off by default, so the test TU isn't in CI's compile_commands.json); it needs a reachable system bus and is intended for local/CTest runs. It skips gracefully with success if no bus is present.
• The one NOLINT(performance-unnecessary-value-param) in the test marks an API-mandated by-value parameter: sdbus deduces reply-handler argument types via function_traits and requires std::optional<sdbus::Error> by value (a const-ref signature fails to compile), matching every existing async handler in the tree.

Next

Step 2 — convert UdevMonitor and InputReader into EventSources and move the three BlueZ clients onto EventLoop, deleting their threads and mutexes. That step first needs the design decision flagged in the scoping: whether synchronous D-Bus calls inside handlers stay synchronous (simplest, but a slow call stalls input reading) or move to async / a dedicated HID read thread.