Adds batch/event shape test and fix failure

[feynmanliang]

NOTE: it may be incorrect to remove event_dim=1 from the bijector; @stefanwebb to confirm

This change is 

[feynmanliang]

The batch_dim erroneously gets reinterpreted into an event_dim when this is 1

[stefanwebb]

Ah, shouldn't this be by design? The AffineAutoregressive makes the final dimension of a random variable dependent, so will introduce correlations if the base distribution's final dimension is a batch dim

[feynmanliang]

Should AffineAutoregressive (or any Bijector) continue across batch dimension? Across batch dimension, samples are independent (https://www.tensorflow.org/probability/api_docs/python/tfp/distributions/Distribution#shapes_2) so I'm inclined to say a completely separate bijector (which cannot use autoregressive predictors from the previous batch instance) should be instantiated for a total of prod(batch_shape) bijectors.

This is similar to sample_dim, except there we re-use the same Bijector due to IID whereas here samples across batch_dim are not identically distributed.