Fix data races in preeditor access across D-Bus/worker threads

[Narcolepsyy]

Problem

IBus dispatches each incoming D-Bus method on its own goroutine (go conn.handleCall in godbus). That means Reset, FocusIn, and FocusOut run concurrently with the keyPressChan worker goroutine (keyPressCapturing). Both paths mutate the shared bamboo preeditor, which is not safe for concurrent use — a data race that can corrupt the buffer and produce wrong output (e.g. oof → ôf instead of ồ, stale words after focus changes).

Changes

• Guard preeditor mutations with the engine mutex in FocusOut, Reset, bsProcessKeyEvent, and keyPressHandler.
• resetBuffer: move the getRawKeyLen() empty-buffer guard inside the lock in both branches. It previously read the preeditor unlocked on the FocusIn → checkWmClass → resetBuffer path, racing the worker goroutine.
• Reset: drain keyPressChan with a non-blocking select/default loop instead of len()>0 + <-. Other goroutines also receive from the channel, so the blocking form could hang the D-Bus thread on an empty channel (TOCTOU).
• SendBackSpace: clamp n to [1, maxBackSpacePerCycle] and reset the fake-backspace counter after each send cycle, so a dropped synthetic backspace can't leave a permanent offset.
• Discard the ephemeral preedit buffer on Reset/FocusOut in PreeditIM only; backspace modes preserve the buffer (apps call Reset aggressively mid-word).

Testing

CGO_ENABLED=1 go test . -mod=vendor -race -run "TestReset|TestFocus|TestFakeBackspace|TestGetOffset|TestConcurrent" -v

• TestConcurrentResetAndKeyPress stresses Reset/FocusIn/FocusOut against the worker goroutine under -race.
• Buffer-semantics tests per input mode (PreeditIM clears, backspace modes preserve), keyPressChan drain, getOffsetRunes, and fake-backspace reset.
• Full suite passes clean under -race.