More consistent Tensorflow chains

[williamjameshandley]

Whilst working on #7, it became clear that we are using multiple versions of normal transformations, tensorflow bijector chains, scaling and uniform distribution quantiles.

This should ideally be more unified, so that at least we are using the same functions for transformations (rather than tensorflow_probability.distribution functions and tensorflow_probability.bijector functions), but even better would be to include the pre-processing _forward_transform and _inverse_transform as part of the full bijector chain.

[htjb]

Okay so been a while since this issue was opened and the code has changed a bit but the _forward_transform and _inverse_transform functions still form an integral part of the code. A more consistent way to incorporate the gaussianization of the data (see figure 2 in this paper) that these functions perform is to define

self.bij = tfb.Chain([tfb.Shift(self.theta_min), 
                              tfb.Scale(self.theta_max-self.theta_min), 
                               tfb.NormalCDF(),
            *[tfb.MaskedAutoregressiveFlow(made) for made in self.mades]])

for the maf. We can then remove the normalisation step for theta in _training I think. This needs testing and I still need to work out what to do with the log probability and sampling but I will try and look into it.