Bnaf Issue

[danielward27]

Fixes #102, by using a LeakyTanh bijection, that defaults to a linear transform above or below max_val. Happy for more investigation/discussions, but this seems a reasonable fix for now.

Breaking changes:

• The activation for BlockAutoregressiveNetwork no longer needs a specific form, and instead can be any (potentially trainable) callable or scalar unconditional bijection.
• Renaming of TanhLinearTails to LeakyTanh could break users code (e.g. if subclassing TanhLinearTails, or explicitly relying on the class name).