Support for outputting embeddings

[joefutrelle]

This pull request adds penultimate-layer embedding output. A one-time ONNX graph surgery exposes the embedding tensor as a second model output, then --embeddings captures it alongside class scores in the same forward pass and writes it per bin.

• add_embedding_output.py (new): auto-detects the pre-head tensor (final Gemm/MatMul input) and adds it to the model's graph outputs; --tensor-name override.
• cli.py: new --embeddings / --embeddings-only / --embeddings-outfile flags; write_embeddings writes Parquet (pid + embedding as fixed_size_list)
• sanstorch.py / withtorch.py: capture outputs[1] and accumulate it alongside scores.
• pyproject.toml: new [embeddings] extra (pyarrow).
• README: install row, options, and an Embeddings section.

Key questions:

• is Parquet the right output format for embeddings?
• is float16 sufficient output precision or is it a premature space optimization?
• are the README changes accurate and sufficient?