data.ERA5_and_Forcing_MultiStep does not behave correctly when train/valid_batch_size is greater than 1

[kanz76]

The current implementation of data.ERA5_and_Forcing_MultiStep samples item based on the forecast_step parameter`, setting the batch_size to greater than 1 can result in a mismatch of input and target tensors.

[yingkaisha]

For now, I will add an assertion in the credit.parser first to encourage the use of batch_size=1.

I will look into this later.