Scaling bugs fixes

[zakv]

There are a few bugs with scalings, which I've patched in this PR.

The first one is that the uncertainties were scaled the wrong way in GuassianProcessLearner. When normalizing costs, the cost_scaler divides them by cost_scaler.scale_ (see here) and when rescaling costs back to their original range it multiplies the scaled cost by cost_scaler.scale_ (see here). The same should be done for the uncertainties, but previously the multiplication and division were switched. This PR corrects that issue. Furthermore the uncertainties in GaussianProcessVisualizer.return_cross_sections() weren't rescaled back, and this has been corrected as well.

The second issue is more minor. The value of cross_parameter_arrays in GaussianProcessLearner.return_cross_sections() and NeuralNetLearner.return_cross_sections() was scaled previously, but does not need to be scaled. When it is created, its values go between self.min_boundary and self.max_boundary which aren't scaled, so there's no scaling to undo. This was likely a small mix-up where the scaling was meant to be applied to uncertainty_arrays but was accidentally applied to cross_parameter_arrays instead, and then that mix-up was copy/pasted to NeuralNetVisualizer.return_cross_sections() . Fortunately the value returned for cross_parameter_arrays isn't used anywhere in M-LOOP, so it didn't cause any other issues.

Changes proposed in this pull request:

• Correct uncertainty scaling in GaussianProcessLearner and GaussianProcessVisualizer
• Correct cross_parameter_arrays scaling in GaussianProcessVisualizer and NeuralNetVisualizer

@qctrl/support @charmasaur

[zakv]

I should also mention that there are a lot of scalings done in neuralnet.py. I haven't tested any of those, but they all looked correct when I read through them!

[charmasaur]

Nice catch!