Optimize speech preview (getByteTimeDomainData + UI update frequency)

[ysdede]

Context

• Speech preview uses AnalyserNode.getByteTimeDomainData + per-frame conversion to Float32Array for waveform bars.
• UI updates occur at ~60fps via onVisualizationUpdate and Waveform rAF loop.

Evidence

• src/lib/audio/AudioEngine.ts → getBarLevels() converts Uint8Array to Float32Array every call.
• src/App.tsx → onVisualizationUpdate calls appStore.setBarLevels(audioEngine.getBarLevels()) every tick.
• src/components/Waveform.tsx draws every rAF, even when not recording.

Impact

• Continuous preview work adds CPU even when transcription is idle; can be a top driver during capture.

Actions

• Throttle preview updates to <=20–30fps (separate from inference tick).
• Skip getByteTimeDomainData when isRecording is false or widget is hidden.
• Consider downsampling the analyser buffer (e.g., 256 → 64) or reusing a single Float32Array without per-frame allocations.
• Optionally compute preview in AudioWorklet and send only min/max to UI.

Acceptance

• Preview CPU drops measurably with recording active but model idle.
• Trace shows reduced main-thread activity during capture when only preview is running.