Add Voxtral Mini 3B contrib model (audio-language)

[jimburtoft]

Summary

• Adds Mistral AI's Voxtral Mini 3B audio-language model to contrib
• Decomposed pipeline: Whisper-like encoder via torch_neuronx.trace(), Llama LLM via NxDI ImageToTextModelWrapper
• Supports text-only generation, audio transcription, and audio understanding

Model Details

• Source: mistralai/Voxtral-Mini-3B-2507 (Apache 2.0)
• Architecture: 637M audio encoder (32L) + 25M projector + 3.3B Llama LLM (30L)
• Precision: BF16
• Instance: trn2.3xlarge (TP=1) and inf2.xlarge (TP=1)

Performance

Instance	Throughput	TTFT
trn2.3xlarge	58.5 tok/s	418ms
inf2.xlarge	28.4 tok/s (text), 27.4 tok/s (audio)	--

Key Technical Contributions

• Reuses NxDI Llama: VoxtralTextModel extends NeuronLlamaModel -- no custom attention/MLP
• Scatter-based audio injection: Uses NxDI's scatter_by_index_put for audio token embedding
• PixtralInferenceConfig reuse: Voxtral's encoder+projector+LLM maps to Pixtral's vision+LLM config pattern
• Hardware-aware compiler args: Auto-detects trn2 for --lnc=2 flag

Testing

• 9 files: README.md, benchmark_encoder.py, src/{init,modeling_voxtral,utils/init}.py, test/{init,integration/{init,test_model},unit/init}.py
• Integration tests: model compilation, text generation (correctness + determinism), audio transcription
• Validated on trn2.3xlarge and inf2.xlarge with SDK 2.28