Avoid repeated mel normalization/copy for overlapping windows

[ysdede]

Context

• getFeatures() allocates a raw window buffer and then normalizes into a second buffer for every request.
• v3/v4 request overlapping windows, so most frames are re-normalized repeatedly.

Evidence

• src/lib/audio/mel.worker.ts → getFeatures copies into raw then normalizeMelFeatures() allocates out.
• src/lib/audio/mel-math.ts → normalizeMelFeatures allocates a new Float32Array per call.

Impact

• Extra allocations + CPU in feature extraction under continuous streaming.

Actions

• Cache normalized frames per mel-bin in the circular buffer, or normalize in-place when raw is no longer needed.
• For overlapping windows, reuse cached normalized frames and only normalize new frames.
• Add a small LRU cache of recent getFeatures windows keyed by (start,end,normalize).

Acceptance

• Feature extraction CPU time drops (p95) in a 30s capture.
• Worker GC time reduced vs baseline.