era5 encoding is lost with preprocess_ERA5

[veenstrajelmer]

When reducing on non-decoded dataset (decode_cf=False upon opening), this preserves the encoding attrs:

ds_reduced = ds_raw.max(dim="expver", keep_attrs=True)

However, in preprocess_ERA5, the input dataset is already decoded and this does not help in preserving encoding (it will be dropped by the reduction). This means the reduction in this function prevents dtype int in a to_netcdf file already. Therefore, prevent_dtype_int is of no value in this function if this expver dimension is present. These files will then be written unzipped to float32 dtypes. It would be good to make this more robust and transparent.

[veenstrajelmer]

This is not an issue anymore with datasets downloaded from the new CDS-beta as the variable dtypes are now zipped float32 instead of scaled int. Therefore this issue is not relevant anymore, only for old datasets.

Furthermore, the prevent_dtype_int() function was also simplified and merged with dfmt.preprocess_ERA5() so this issue is not that relevant anymore. More info in #942