Simplify stitching by retaining prediction objects

[MartinSJRogers]

This PR builds upon #19 by retaining the patchwise predictions as deepsensor.prediction objects. This reduces the indentation and number of for loops within stitch_clipped_predictions(). Because the code uses the .where() function to merge patchwise predictions, there is still a necessity to generate a blank prediction object with the same extent, keys and data variables at the entire extent of X_t. There may be a shorter way to do this using prediction.py but I could not see this.

[davidwyld]

Note: I've just tried to review the last commit (58e9076) because the others seem to have been on the branch merged in #19, I'm not sure why the commits are on both branches.

Seems to do the trick! Nice work!

[davidwyld]

I'm a little wary of this line, is this getting the first item in patch_pred? If so, I think that using next(iter( is a bit clunky, but also because dictionaries aren't guaranteed to be ordered consistently. Could we do this with a particular key? first_value isn't used so it seems like we only need the key here.

[davidwyld]

Looking more into this, apparently dictionaries are ordered for python 3.7+ so we're safe. But it might be safer to use something like list(patch_pred.keys)[0] if we only need the key.

[MartinSJRogers]

Happy to change it to this. There's always many ways to skin a cat, but I agree your way seems safer.

[MartinSJRogers]

The reason for doing this is just because I only need one variable name to work out things like coordinates and extent. So actually it shouldn't matter in this instance if the variable names are different, but good to ensure its consistent just in case.