More flexibility for broadcasting shapes in `DensityEstimator`s

[michaeldeistler]

As raised in #1066 the DensityEstimators currently have the following limitations:

Not all DensityEstimators can...

• ...handle batch_shapes (but only batch_dims, i.e. scalar values)
• ...evaluate one datapoint under multiple conditions with log_prob without adapting the batch_dim of the datapoint
• ...perform log_prob on data without iid_dim_input. I.e. log_prob((50,3), (50,4)) might fail, but log_prob((1,50,3), (50,4)) will work