Add WeatherNext-style ONNX conversion demo

[Copilot]

WeatherNext uses a JAX/Haiku + xarray architecture that does not map to the existing HuggingFace model registry flow. This PR adds a runnable ONNX demo that sketches the WeatherNext one-step forecast contract and graph workflow under examples/.

• Demo workflow

  • Adds examples/weathernext.py.
  • Builds a small deterministic grid→mesh→grid forecast graph.
  • Saves the result as a standard Mobius ModelPackage.
  • Supports optional ONNX Runtime validation with --validate.

• WeatherNext-style graph contract

  • Inputs:
    • input_state: weather variables on a lat/lon grid
    • forcings: known future forcing variables
    • sample_noise: stochastic FGN-style noise channels
  • Output:
    • next_state: one-step forecast grid

• Coverage

  • Adds tests/weathernext_example_test.py to verify the exported ONNX graph I/O contract.

Example:

PYTHONPATH=src python examples/weathernext.py output/weathernext-mini --validate

• Fixes Convert weather next #468

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[justinchuby]

@copilot what I wanted was a formal support of this model in the modeling/integation files, and a demo script that can run this model on real data.

[Copilot]

@copilot what I wanted was a formal support of this model in the modeling/integation files, and a demo script that can run this model on real data.

Implemented in f588c1a (with the WeatherNext support commits immediately before it): the WeatherNext graph is now formalized under mobius config/model/task/integration modules, and examples/weathernext.py uses that path with NPZ/xarray real-data inputs plus optional converted weights.