fix: seed regional decoder observation nodes with encoder features

[AswaniSahoo]

Pull Request

Description

RegionalForecaster seeds its decoder observation nodes with torch.zeros, discarding the per-observation features the encoder just computed - the same bug that #237 fixed in StretchedForecaster. Each observation's predicted change can only be read back from its assigned cell's pooled, message-passed feature, so the observation's own values reach the output only through the fixed residual. In #237 this collapsed the stretched model to the persistence baseline on a 2-sample overfit (floor 0.173 = persistence, dropping to 0.136 once fixed).

This applies the identical fix here: seed the decoder's observation nodes with the encoder's per-observation features (nodes[:num_obs]) instead of zeros, so the learnable delta can specialise per observation. No new parameters.

Follow-up to #237, which flagged this pattern in RegionalForecaster. Related to #3

How Has This Been Tested?

Added test_decoder_seeded_with_encoder_obs_features to tests/test_regional_forecast.py - a plain pytest that hooks the decoder GNN, runs a forward pass, and asserts the observation nodes it receives are non-zero (they carry the encoder output rather than the old zeros seed). It fails on the previous code.

Commands:

• pytest tests/test_regional_forecast.py -q -> 14 passed
• ruff check graph_weather/models/regional_forecast.py tests/test_regional_forecast.py -> clean

Checklist:

• My code follows OCF's coding style guidelines
• I have performed a self-review of my own code
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• I have checked my code and corrected any misspellings