Fix probabilistic variance underflow

[Panchadip-128]

Describe your changes

This PR resolves two critical numerical and logical flaws in the probabilistic forecasting (--output_std) engine:

1. Fixed softplus Underflow (NaN crashes): In neural_lam/models/base_graph_model.py, the standard deviation calculation is now clamped to a minimum of 1e-6. This prevents the network from producing a machine-zero variance which previously caused division-by-zero errors in NLL and CRPS metrics, leading to irreversible NaN training losses.
2. Corrected Ensemble Sampling: The ARModel._sample_ensemble method in neural_lam/models/ar_model.py was previously discarding the model's predicted uncertainty map in favor of a hardcoded 0.01 noise fallback. This has been refactored to utilize the model's dynamically predicted pred_std for physically grounded ensemble generation.
3. Regression Testing: Added test_base_graph_model_prevents_softplus_underflow_nans and test_ar_model_ensemble_samples_from_pred_std to tests/test_probabilistic_forecasting.py to assert that these mathematical stability and logic requirements are met.

Motivation and Context: These bugs made probabilistic training unstable and produced deceptively uniform ensemble spreads that ignored the model's internal confidence.

Dependencies: No new dependencies.

Issue Link

solves #526

Type of change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 💥 Breaking change (fix or feature that would cause existing functionality to not work as expected)
• 📖 Documentation (Addition or improvements to documentation)

Checklist before requesting a review

• My branch is up-to-date with the target branch (upstream/main rebased and force-pushed).
• I have performed a self-review of my code.
• For any new/modified functions/classes I have added docstrings.
• I have placed in-line comments to clarify intent.
• I have updated the README (Not applicable as these are internal bug fixes).
• I have added tests that prove my fix is effective.
• I have given the PR a name that clearly describes the change.

Author checklist after completed review

• I have added a line to the CHANGELOG.md describing this change.
  • fixes: softplus variance underflow and ensemble sampling logic.

[ronilmitra7]

Nice catch on the NaN crashes! I’ve definitely been there with PINNs and seen how one zero variance can just blow up a whole training run. Clamping at 1e-6 is a solid move for keeping the NLL/CRPS stable.

Also, good to see that 0.01 noise replaced with the actual pred_std in the ensemble logic. I’m curious, did you run into these NaN crashes mostly during the initial epochs or during longer auto-regressive rollouts?