feat: add MLX inference for Apple Silicon (CTC + RNNT)

[misteral]

Summary

Native MLX inference for GigaAM v3 on Apple Silicon — supports both CTC (139× realtime) and RNNT (48× realtime, ~9% lower WER) models.

Supersedes #61 (cleaned up: rebased on main, removed dev artifacts).

What's included

File	Description
mlx_convert/gigaam_mlx.py	Full MLX model: Conformer encoder (16 layers, 768d, RoPE), CTC + RNNT heads, mel spectrogram, streaming
mlx_convert/convert_gigaam_to_mlx.py	PyTorch → MLX conversion (safetensors + config.json)
mlx_convert/gigaam-cli	Single-file transcription CLI
mlx_convert/gigaam-stream	Real-time streaming (live mic + file)
mlx_convert/gigaam-transcribe	Shell wrapper
mlx_convert/README.md	Documentation: Python API, CLI, benchmarks

Architecture

Audio (16kHz) → Log-Mel (64 bins) → Conv1d Subsampling (4×)
  → 16× Conformer (RoPE MHSA + GLU Conv + SiLU FFN)
  → CTC Head → Greedy Decode
    or
  → RNNT Head (Joint + LSTM Decoder) → Greedy Decode

Key details:

• RoPE applied before Q/K/V projections (matching original PyTorch model)
• Mel filterbank saved from PyTorch (exact match, no recomputation drift)
• All Conv1d weights transposed: [out, in, K] → [out, K, in] for MLX
• RNNT LSTM weights properly mapped to MLX layout

Performance (Apple M4)

Model	Batch (11s audio)	Realtime factor	Model size (fp16)
CTC	81ms	139×	421 MB
RNNT	230ms	48×	423 MB

Python API

from gigaam_mlx import load_model, load_audio

# CTC (fastest)
model = load_model("./gigaam-v3-ctc-mlx")
text = model.transcribe(load_audio("audio.wav"))

# RNNT (higher quality)
model = load_model("./gigaam-v3-rnnt-mlx")
text = model.transcribe(load_audio("audio.wav"))

# Streaming
for r in model.stream_generate(load_audio("audio.wav")):
    print(r.cumulative_text)

Testing

Tested on Apple M4 with various Russian speech samples. Output matches PyTorch reference (character-level exact match on short utterances).