Inconsistent behavior with log_prob on transformed distributions

The same input in a transformed log_prob will raise an exception, or provide the correct value, depending on the sequence of execution.

Talking with @brianwa84 the jacobian in the bijector might be expecting the wrong data type, if I understood correctly.

Below is the example that him and I talked about in person

dist =  distributions.Bernoulli(probs=.5, dtype=tf.int32)

_bijector = bijectors.Identity()
transformed_dist = distributions.TransformedDistribution(
            distribution=dist, bijector=bijectors.Invert(_bijector)
        )

To reproduce run the above code interactively, then run each line below in sequence.

log_prob(int) will raise an exception until

transformed_dist.log_prob(1)

log_prob(float) is executed

transformed_dist.log_prob(.5)

Then log_prob(int) will provide the correct value

transformed_dist.log_prob(1)

Start by running the provided Bernoulli, Identity, Invert, and TransformedDistribution example interactively, then compare the three log_prob calls in sequence. Investigate the transformed-distribution or bijector Jacobian path for the dtype-dependent behavior; done means integer log_prob behaves consistently regardless of whether the float call ran first.