Major update for the inference routine + QoL improvements

[yingkaisha]

New features

• Now grad_max_norm: 0 will turn off gradient clipping

• In credit.parser, conf['trainer']['mode'] is verified to be one of the fsdp, ddp, or none.

• In credit.parser, the data availability checks now produce warnings instead of errors.

  • If you require to predict 2024 but only have data up to 2023, you will receive warnings, and the inference routine will stop at the end of 2023.

• The following config options will now trigger warnings:

  • FuXi with use_spectral_norm: False and mode: fsdp --- this does not work. The reason is unknown.

• Datasets defined in rollout_to_netcdf.py and rollout_metrics.py are now combined as Predict_Dataset_Iter in credit.data

Bugfix

• Diagnostic variable issues in Bugfix on the production of diagnostic variables in rollout_to_netcdf.py + credit.postblock update + other minor bugfix #120 is fixed in rollout_metrics.py.
• Bugfix the model weights loading in rollout_to_netcdf.py and rollout_metrics.py for ddp and single-GPU mode.

[yingkaisha]

Done.

Some files are touched because of the ruff reformat.