Integrate internal feedback

[bwdGitHub]

I've modified the example following internal feedback. This version should more closely resemble other documentation examples in style and content.

Some notable changes:

1. The targets are no longer normalized. This means the training losses are on a different scale to before. This seemed to be OK - the final results still look good. Inputs are normalized via the image3dInputLayer defaults.
2. Use of combinations over doing this manually to create all combinations of the varying parameters.
3. Use of fillmissing over manually finding NaN and replacing it.
4. Only solve the PDE for training data at the final time. This drastically reduces the size of the results objects since you're only storing the solution at the final time - from about 2GB to 400MB.

There's a couple extra changes I'll try push up on this branch:

1. Use bounds over finding min and max independently.
2. Swap from meshgrid to ndgrid This saves some awkward permute-s in the griddedInterpolant call used at the test/inference section at the end.

[bwdGitHub]

I've addressed the extra changes 1. and 2. in the latest commit:

1. Use bounds - simplifies the code over separate min and max calls.
2. Use ndgrid rather than meshgrid, integrates better with griddedInterpolant later on.