Fix rounding errors in DeepSensorModel.predict coordinates from normalise-unnormalise operations

[polpel]

This adds a unit test which checks that normalising and unnormalising an xarray object with DataProcessor (as is done e.g. in model.predict) should return the exact same coordinates as the original object.

This test currently fails, as pointed out in Issue #19

[tom-andersson]

Hi @polpel, thanks very much for opening this. It's important to fix this to avoid the silent dropping of data in #19.

The reason for the slightly differing coordinate values is simply floating point rounding error / numerical noise. I tried moving to float64s and while this reduced the magnitude of the rounding errors, they were still present, and it looks like xarray needs the coordinates to align exactly for operations like subtraction (as mentioned in #19), which has been discussed on the xarray repo (e.g. pydata/xarray#2217, pydata/xarray#6573).

Since we can't prevent rounding errors, the problem is actually with the way the high-level DeepSensorModel.predict method normalises the X_t coords for prediction and then unnormalises this to return to the user. I've fixed this now with 74cc5b8 and moved/refactored your unit test into tests/test_model.py.

Tests are passing, so I will merge with main. This closes #19.

[tom-andersson]

Note for future reference that this is missing an indentation level @polpel - it should be part of the test class.