perf(voxtral): O(1)-per-chunk conv-stem for streaming (no more O(n²) prefix recompute)

[beshkenadze]

Summary

Follow-up to #16. VoxtralRealtimeStreamSession recomputed mel + conv stem over the entire accumulated audio on every chunk (O(n²) per stream, plus a CPU reflect-pad round-trip of the whole buffer each step). Mid-stream chunks now derive padded-audio geometry arithmetically and compute only the new conv rows from a bounded raw-sample window — O(1) per chunk regardless of stream length.

Why it's bit-exact

• conv row f needs mel frames [2f-3, 2f+1]; a mel frame covers [m·hop − win/2, m·hop + win/2) of the left-padded audio — bounded lookback/lookahead.
• The STFT reflect pad resolves to zeros (the streaming left pad is ≥ win/2 zeros), and frozen frames never reach the right zero-pad (the dropped-partial-frame token formula guarantees ≥ 40 samples of slack for any chunk size).
• First chunk and finish() keep the full stem, so the final transcript stays equal to offline generate().

Verification

WER(stream vs offline) = 0.0000 on:

• conversational_a at 90 / 333 / 480 / 730 ms chunk sizes (incl. non-token-aligned)
• intention (1.5s), conversational_fr
• a 12-minute stream (1520 chunks): median step 0.193s, p90 0.334s @ 480ms budget, end-to-end realtime — step time no longer grows with stream length

swift build + swift build --build-tests clean.