Realnvp/Nice odd dim

[htjb]

Description

Coupling layer flows have been failing for odd numbers of dimensions. In the coupling models we define a pass_size = in_size // 2 where in_size is the number of dimensions and a net_in_size = in_size - pass_size. Then in each layer we split the data in to xa = x[:, :pass_size] and xb = x[:, pass_size:] so xb has shape (len(samples), net_in_size) and xa has shape (len(samples), pass_size). xb is then passed through a neural network that outputs values (s and m in a realnvp) with shapes given by net_in_size. Then when you operate with these outputs on xa you get shape mismatches e.g. in a realnvp you do xa = xa * jnp.exp(s) + m. This works fine for even dimensions because pass_size = net_in_size.

Checklist:

• I have performed a self-review of my own code
• New and existing unit tests pass locally with my changes (python -m pytest)
• I have added tests that prove my fix is effective or that my feature works
• I have appropriately incremented the semantic version number in both README.rst and margarine/_version.py