Add AutoML via Optuna

[crwhite14]

This pull request adds Optuna for Darcy flow, meant as an example so that users can set up automatic hyperparameter tuning for their models and use-cases.

Changes

• The evaluate method in Trainer now returns the errors dict
• New file, hpo/tune_darcy.py, that launches optuna. It is similar to train_darcy.py, but most of the contents are inside a method, objective(trial). At the start of each trial, Optuna samples the most promising hyperparameters based on the results of the previous trials.

How to use Optuna

• Specify the hyperparameters to search over, and their ranges, here: https://github.com/crwhite14/neuraloperator/blob/b768ce2440ad9cd13338ec5071a9da6f4fc077e6/scripts/hpo/tune_darcy.py#L88
• Specify the objective to search over here (typically validation loss): https://github.com/crwhite14/neuraloperator/blob/b768ce2440ad9cd13338ec5071a9da6f4fc077e6/scripts/hpo/tune_darcy.py#L207
• Then run python hpo/tune_darcy.py

[JeanKossaifi]

Thanks @crwhite14 it looks great!