Tensor dimension and order issues in data workflow

[yingkaisha]

• In credit.data, the ERA5_and_Forcing_Dataset class produces xarray.Dataset with dimensions of [time, level, latitude, longitude]
• In credit.transforms, class ToTensor_ERA5_and_Forcing collects xarray.Dataset variables and produces tensors of [time, var, level, lat, lon]
• In credit.trainers, the Trainer class in trainerERA5_v2.py collects [batch, time, var, level, lat, lon] from the dataloader and permute it to [batch, var, time, level, lat, lon]

The entire process above can be handled better with ToTensor_ERA5_and_Forcing produces [var, time, level, lat, lon]. This will avoid the tensor permutation in trainerERA5_v2.py.

Also note that all current models (their embedding layers) take [batch, var, time, lat, lon] tensors with "channel-first" and patch_size = [patch_time, patch_lat, patch_lon]. The improved data workflow needs to be compatible with this model requirement.