feat: OpenVINO CPU MetricGAN mask; 100 ms hop

Per-worker OpenVINO CPU runs the MetricGAN BLSTM mask; SpeechBrain
keeps STFT/ISTFT. ENHANCE_INFER=auto with torch fallback. Hop is
0.1 s after isolated last_ms ~6 ms (model+ov+beamform).

README.md

@@ -101,22 +101,25 @@ HDMI-A-7 is replaced by the grid).
 
 ## Audio cleanup
 
-C920 mics are 16 kHz **stereo**. Each hop (default 200 ms) is kept
+C920 mics are 16 kHz **stereo**. Each hop (default 100 ms) is kept
 stereo through a SpeechBrain preprocessor, then published as mono:
 
 1. Delay-and-Sum beamform (GCC-PHAT TDOA, frozen for
    `BEAMFORM_TDOA_EVERY=5` hops) — live path from the SpeechBrain
    [multi-microphone beamforming tutorial](https://speechbrain.readthedocs.io/en/latest/tutorials/preprocessing/multi-microphone-beamforming.html)
 2. `speechbrain/metricgan-plus-voicebank` neural enhancement on a 1 s
-   rolling window (`ENHANCE_CHUNK_S=1.0`, `ENHANCE_HOP_S=0.2`)
+   rolling window (`ENHANCE_CHUNK_S=1.0`, `ENHANCE_HOP_S=0.1`).
+   The MetricGAN **mask** runs on OpenVINO CPU (`ENHANCE_INFER=auto`);
+   STFT/ISTFT stay in Torch. `ENHANCE_INFER=torch` forces SpeechBrain
+   `enhance_batch`. Do not use the Arc B580 for this net.
 
 The 1 s window is **past context**, not +1 s delay: only the latest hop
 is published. Algorithmic delay ≈ hop + compute.
 
-Each worker caps Torch/BLAS to `TORCH_NUM_THREADS=1` (default) so 14–20
-processes do not oversubscribe the i7-14700. This host has an Intel Arc
-B580; the venv Torch is NVIDIA CUDA and **cannot** use it. Do not set
-`run_opts={"device":"cuda"}`.
+Each worker caps Torch/BLAS/OpenVINO to `TORCH_NUM_THREADS=1` (default)
+so 14–20 processes do not oversubscribe the i7-14700. This host has an
+Intel Arc B580; the venv Torch is NVIDIA CUDA and **cannot** use it. Do
+not set `run_opts={"device":"cuda"}`.
 
 One worker process per camera loads the modules once and reuses them.
 `enhance_batch` must be called with `lengths=torch.tensor([1.0])`