[ONNX] Add Chronos-2 covariate export validation

[Jameswlepage]

Summary

This builds on and supersedes the initial Chronos-2 ONNX export work in #359.

It updates the Chronos-2 ONNX export path so users can export a real Chronos-2 model with future covariates, repair the exported ONNX graph, and validate ONNX Runtime output against PyTorch.

Changes

• Export amazon/chronos-2 by default with future covariates enabled.
• Use the legacy TorchScript ONNX exporter explicitly with dynamo=False for compatibility with current PyTorch defaults.
• Keep batch size dynamic while documenting fixed context and forecast lengths.
• Run the ONNX graph repair pass automatically after raw export.
• Repair fix_onnx_model.py so it fixes Gather index dtype issues without incorrectly advertising a dynamic horizon.
• Add a standalone PyTorch-vs-ONNX parity harness for grouped batches, future covariates, and missing values.
• Update quantized-model validation to infer the exported fixed shapes from the ONNX graph.
• Add ONNX export documentation with supported inputs, outputs, and limitations.
• Adjust Chronos patching and missing-value normalization paths so the Chronos-2 model exports cleanly.

ONNX Contract

The default export has:

• context: float32[batch_size, 512]
• group_ids: int64[batch_size]
• attention_mask: float32[batch_size, 512]
• future_covariates: float32[batch_size, 64]
• num_output_patches: int64[] when exposed by the exporter
• quantile_preds: float32[batch_size, 21, 64]

Batch size is dynamic. Context length, future covariate length, and prediction length are fixed by the traced export.

Validation

Local checks run:

• python -m py_compile scripts/onnx/export_chronos2_to_onnx.py scripts/onnx/fix_onnx_model.py scripts/onnx/quantize_chronos2.py scripts/onnx/validate_chronos2_onnx.py src/chronos/chronos_bolt.py
• python scripts/onnx/export_chronos2_to_onnx.py --model_id test/dummy-chronos2-model --output_dir ... --device cpu --validate
• python scripts/onnx/validate_chronos2_onnx.py --model_id test/dummy-chronos2-model --onnx_path .../model.onnx --context_length 512 --num_output_patches 4
• python scripts/onnx/quantize_chronos2.py --input .../model.onnx --output .../model_quantized.onnx --mode dynamic --validate

A real amazon/chronos-2 ONNX covariate export and parity reports are published here:

https://huggingface.co/TSFM-ai/chronos-2-onnx

The real export parity report covers batch sizes 1 through 5, shared and distinct group_ids, several future-covariate patterns, missing context values, and missing future-covariate values.

[Jameswlepage]

@kashif I opened this as a follow-up to #359 with the Chronos-2 covariate export fixes, ONNX graph repair flow, and PyTorch-vs-ONNX parity validation. Would appreciate your review when you have a chance.