Enhance documentation for ScaleMatvecTriL bijector reg. "batch behavior"?

Hi,

porting the new bijectors to R, I was a bit confused by ScaleMatvecTriLs behavior (as opposed to what I was expecting from the documentation,

https://github.com/tensorflow/probability/blob/557351a2b79e4bd5a9025e791c81b612ab17bf38/tensorflow_probability/python/bijectors/scale_matvec_tril.py#L42

My test case was like this

tril = np.array([[1., 0],[1, 1]])
x = np.array([[1., 2], [3, 4]])
tril @ x

#array([[1., 2.],
#          [4., 6.]])


bijector = tfp.bijectors.ScaleMatvecTriL(tril)
bijector.forward(x) 

#<tf.Tensor: shape=(2, 2), dtype=float64, numpy=
#array([[1., 3.],
#          [3., 7.]])> 

From looking at

https://www.tensorflow.org/api_docs/python/tf/linalg/LinearOperatorLowerTriangular

, I think what is going on is that the operator acts on every row separately (treating the first dimension as a batch dimension).

Do you think it makes sense to go into a bit more detail on this in the documentation, to make it easier on users to understand the behavior?

Thanks!

Start with tensorflow_probability/python/bijectors/scale_matvec_tril.py at the documentation referenced near line 42, then compare its explanation with tf.linalg.LinearOperatorLowerTriangular and the reported ScaleMatvecTriL example. Clarify how matrix dimensions are interpreted and how batch behavior produces the shown result; done means the documentation explains this behavior unambiguously.