The same dropout value is used for all NaNs

[dfulu]

The implementation of the configurable dropout value is problematic or misleading. In the config, it states that it is used to replace values that are dropped out, but in reality is it used to replace any NaN values in the data. This could be NaNs that are in the raw zarr - which is possible with satellite data at least.

We should define what the desired behaviour is and make this NaN filling more clear

[dfulu]

This will clash with changes which are part of #400, so on hold for now