feat(nemotron_asr): incremental NemotronASRStreamSession for live mic

[beshkenadze]

What

Adds NemotronASRStreamSession — a true incremental (online) streaming session for Nemotron 3.5 ASR, so a live caller (e.g. a mic feeding 80–480 ms chunks) gets text as audio arrives instead of only after the whole buffer is in.

let session = model.makeStreamSession(language: "ru")   // or chunkMs: 320
let delta = session.step(samples)   // text decoded so far this call
let tail  = session.finish()        // flush trailing partial chunk

Why

generateStream(...) computes the mel of the whole utterance up front, then walks the cache‑aware encoder — so nothing comes out until the end. On a 47 s clip that's TTFT ≈ 2.49 s ≈ wall. The session emits per frozen chunk: TTFT 2.49 s → 0.09 s (×27), RTF unchanged (~0.048). This is what makes Nemotron usable for live mic / two‑tier STT.

How it stays bit‑identical to generateStream(wholeAudio)

• The preprocessor is normalize: "NA" → each mel frame is an independent function of a fixed sample window (no per‑utterance mean/std that shifts as audio grows); preemph is causal.
• The STFT centers with nFft/2 zero‑pad, so mel frame m covers samples [m·hop − nFft/2, m·hop + nFft/2). Frame m is frozen (unaffected by future audio) once m·hop + nFft/2 ≤ bufferLen. The session feeds the encoder only frozen, whole chunks; the trailing partial chunk is flushed in finish(), reproducing the offline right‑pad exactly.
• Encoder + greedy RNN‑T state are lifted into NemotronASRStreamEncoderState / NemotronASRStreamRNNTState; the chunk loop and the RNN‑T decode are now shared by generateStream and the session (SSOT — no duplicated loops; generateStream's observable behavior is unchanged).

Verification

• Parity (real model, mlx-community/nemotron-3.5-asr-streaming-0.6b-8bit, FLEURS‑ru): 31/31 — session(fed @80 ms) == session(fed @480 ms) == generateStream(wholeAudio), byte‑identical transcript.
• TTFT 2.49 s → 0.09 s on a 47 s meeting clip.
• Unit tests on a tiny synthetic model assert session(chunked) == generateStream(whole) and feed‑granularity invariance (gated behind MLXAUDIO_ENABLE_MLX_RUNTIME_TESTS=1, like the other Nemotron MLX tests).

Files to review (5)

• Sources/MLXAudioSTT/Models/NemotronASR/NemotronASRStreamSession.swift — new session + shared RNN‑T decode + makeStreamSession
• Sources/MLXAudioSTT/Models/NemotronASR/NemotronASRStreaming.swift — resumable streamEncodeChunks (state lifted out of the loop)
• Sources/MLXAudioSTT/Models/NemotronASR/NemotronASRModel.swift — generateStream now reuses the shared loops
• Sources/Tools/mlx-audio-swift-stt/App.swift — --stream drives the session for Nemotron (mirrors Voxtral)
• Tests/MLXAudioSTTTests.swift — parity + granularity‑invariance tests

⚠️ Branch base note

This branch is based on upstream/main (which has the NemotronASR base from Blaizzy#195–Blaizzy#198). Your fork's main predates that, so merging here also carries the upstream additions your main lacks (Whisper/Kokoro/Irodori/Nemotron). For a clean session‑only diff, Sync fork → upstream first, then this PR re‑computes to just the 5 files above. Your fork‑only Voxtral incremental‑streaming work is untouched by this branch.

Note: the dev parity/TTFT exe (nemo-parity) is kept off this PR (absolute‑path manifest); it lives on local branch dev/nemo-parity-harness.

[beshkenadze]

Superseded by #21 (clean branch off current main, App.swift demo factored out).