feat: add senselab VAD, eGeMAPS, and embedding processors (#566)

[Chelsy-AI]

Closes #566

Summary

Adds three FeatureProcessor subclasses wrapping senselab for voice/speech analysis in pyhealth/processors/senselab_processor.py:

• SenselabVADProcessor — removes non-speech segments, returns speech-only waveform tensor
• SenselabEGeMAPSProcessor — extracts 88 hand-crafted eGeMAPS acoustic features
• SenselabEmbeddingProcessor — extracts deep neural embeddings via WavLM or ECAPA-TDNN

All three follow the existing AudioProcessor pattern: lazy imports, ImportError with install hint, process() takes a file path, returns a torch.Tensor. Processors are registered in the processor registry and exported from __init__.py.

Tests are in tests/processors/test_senselab_processor.py using pytest.importorskip("senselab") so they skip gracefully in environments where senselab is unavailable.

Note on dev environment

senselab>=1.3.0 requires torch>=2.8.0 with CUDA libraries (~3GB), which conflicts with available disk space on GitHub Codespaces alongside PyHealth's own dependencies. Local macOS testing is blocked by llvmlite 0.47 requiring LLVM 20 while Homebrew ships LLVM 22. Happy to test in a larger environment if maintainers can advise on the recommended dev setup for audio dependencies.

[jhnwu3]

Hey, do you mind giving more context for what this is for?