[Feature request] Graph capturing for ONNX and StableHLO

It would be useful if Candle could capture a graph representation an implemented model using procedural macros. This graph representation can then be used to export the model in the ONNX format so that models trained with Candle can be used with other machine learning frameworks.

It also would allow to transform the graph to StableHLO and then use e. g. IREE to optimize and execute the graph. This might be interesting because it gives access to all backends implemented in the chosen compiler (e.g. AMD's ROCm) and the optimized graph can potentially be faster.

No files, tests, or entry points are named. Start by scoping how procedural macros could capture implemented Candle models and how the resulting graph should be exported to ONNX or transformed to StableHLO for IREE; done requires an agreed design and support for both graph outputs.

Written by the indexing model from the issue text.