Add torch.clamp to trainers + add drop_path in FuXi + bugfix FuXi when use_spectral_norm: False + minor updates on credit.parser

[yingkaisha]

New feature

• data_clamp keyword is added to the data section (e.g., data_clamp: [-16, 16]). CREDIT will perform:

x = torch.clamp(x, min=clamp_min, max=clamp_max)
y = torch.clamp(y, min=clamp_min, max=clamp_max)

During its training and rollout. This will reduce the dmg caused by outliers

• Added proj_drop, attn_drop, and drop_path in FuXi
• A few lines were added in credit.parser to create separated entries for pressure-level and model-level inputs. It does not interrupt any existing functionality on pressure-level datasets.

Bugfix

• Fixed device error on loading FuXi when use_spectral_norm: False

---

• This PR can be verified as follows:

python applications/train.py -c config/example_physics_single.yml
python applications/rollout_to_netcdf.py -c config/example_physics_single.yml

[djgagne]

The changes look good to me.