Auto-differentiable numerical inverse & efficient materialization of BNAF masks

[noahewolfe]

Here, we make two updates which are particularly relevant for block-neural autoregressive flows:

1. Auto-differentiable numerical inverse via the implicit function theorem. (Fixes Autodiff problem with block_neural_autoregressive_flow #176)

• Using the implicit function theorem (I followed https://arxiv.org/abs/2111.00254 in particular), we define a custom jacobian-vector product for NumericalInverse transforms.
• Adds lineax as a dependency for memory-efficient computation of the JVP.
• I've tested that this works using the block_neural_autoregressive_flow and the default greedy bisection search (in a publication to be on the arxiv in the next week or two).
• I added very light unit tests, which pass.

2. Efficient materialization of block masks with Kronecker products

• I found that it was slow to initialize (e.g., just put into memory---no training or anything) block-neural autoregressive flows with large nn_block_dim.
• I traced this to block_diag_mask and block_tril_mask, and reimplemented these with kronecker products.
• These changes pass the unit tests in test_masks.py

Let me know what you think, happy to answer any questions and take any feedback!

[danielward27]

Nice! Cheers. I'm a bit busy at the moment but will try to check it out before the end of the week.

[danielward27]

Thank you so much again for the contribution! It looks good to me, I've just left some comments with some minor suggestions, let me know what you think.

[danielward27]

This change seems to be unused?

[danielward27]

I think we can remove the raise_old_error. If we don't bother using the legacy behavior for a deprecation cycle then we should probably just not bother including it to simplify the code a bit.

[danielward27]

To me it would feel less confusing to replace diffable_inverter argument with use_implicit_diff or use_implicit_differentiation as a key word only argument defaulting to True.