Loading data takes long due to xESFM weight generation on the fly

[jbusecke]

This code block takes very long since it calculates the weights for regridding in the horizontal on the fly.

I think we should save those weights out and provide an option to load the weights from a bucket.

url_weights = ...
ds = preprocess(..., atmos_regridding_weights=url_weights) # reads weights from store
ds = preprocess(..., atmos_regridding_weights=None) # creates weights from scratch

But should we do this as default or not?