TTS speak() blocks the main thread on every spoken selection, freezing the gaze cursor

Why

VocableTextToSpeech.speak() performs blocking binder IPC into the TTS engine process on the caller's thread — setLanguage() (up to twice, with locale fallback) plus the per-call tts.voices enumeration that powers #632's live default-voice fallback — and every production call site invokes it from the main thread. The calls can take tens to hundreds of milliseconds, during which the Choreographer stops ticking: the gaze cursor visibly freezes and gaze input effectively pauses on every spoken selection. This is currently masked by GazeClickable's post-selection hold (the button is deliberately frozen green during the same window), which is how it went unnoticed — it was exposed during #678's smoothing work when that hold was temporarily disabled.

Scope

Move speak()'s blocking engine work off the main thread without changing spoken behavior.

Acceptance Criteria

• speak() performs no blocking TTS engine calls (setLanguage, voices enumeration, speak) on the caller's thread
• Engine work is serialized so concurrent speak() calls cannot interleave engine-state mutations (single dedicated thread, not a pool)
• The stale-selection signal still clears the persisted voice name at all call sites (return value becomes a callback or equivalent)
• Spoken behavior is unchanged: utterance ordering, QUEUE_FLUSH semantics, and isSpeakingFlow transitions
• Verified on-device: the gaze cursor keeps moving through the selection-to-speech gap

Out of Scope

• Whether the post-selection hold itself should remain coupled to TTS completion (a confirmed iOS/Android behavior difference — product question, not yet tracked)
• Any debug flag for skipping the selection hold (not yet tracked)

Links

• https://github.com/willowtreeapps/vocable-android/issues/678 (where the stall was discovered; its work-log documents the smoothing pipeline this stall interrupts)