Eliminate per-window allocations in ring buffer reads

[ysdede]

Context

• RingBuffer.read() allocates a new Float32Array for every window/segment read.
• v3/v2 and segment extraction call read() frequently (windowing + segment verification).

Evidence

• src/lib/audio/RingBuffer.ts (read allocates)
• src/lib/audio/AudioEngine.ts uses ringBuffer.read() in processWindowCallbacks and segment extraction.
• src/App.tsx v2/v3 paths read full windows for transcription.

Impact

• Per-window allocations increase GC pressure and CPU during capture.

Actions

• Prefer readInto() with a reused buffer for window reads.
• For window streaming, consider a SharedArrayBuffer ring to avoid copies.
• Add metrics to compare allocation counts before/after.

Acceptance

• Allocation rate during recording decreases (observed via Performance/Memory).
• Worker GC time reduced in a 30s capture.