Consistent predict keyword options.

This bump to version 1.5.5 brings the possibility to only limit the prediction to only calculate the mean, with the variance and emulated gradient being now optional. Note that the return is now always a 3-element tuple, with some of those elements set to None if they're not being calculated.

This version also gives you the chance to specialise the training set. While that could already be done by changing the PDF of the parameters for the LHS scheme, you can now add reinforced sampling of same areas of interest.

Finally, the dependency on h5py has now been take out (I think!). At some point, we'll have a sensible way to save and retrieve emulators and metadata, but so far hdf5 is causing more headaches than problems it is solving!