Implement the compute_output_shape method on DenseVariational.

[Frightera]

Related to the issue #1505. dense_variational_v2.py does not have this method so it throws:

NotImplementedError: Exception encountered when calling layer "time_distributed" (type TimeDistributed).
    
Please run in eager mode or implement the `compute_output_shape` method on your layer (DenseVariational).

Defining compute_output_shape method fixes this issue. Please see the gist for the reproducible example, fixed example and unit test result.

[Frightera]

@ColCarroll Thanks for approving. This should still be an issue with TFP 0.16 as

probability/tensorflow_probability/python/layers/dense_variational_v2.py

Line 22 in 5172cc7

class DenseVariational(tf.keras.layers.Layer):

does not have compute_output_shape method. Old layers such as DenseFlipout etc has this method:

probability/tensorflow_probability/python/layers/dense_variational.py

Line 189 in 5172cc7

def compute_output_shape(self, input_shape):

[ColCarroll]

Thanks for the contribution, @Frightera! I'm not sure I understand your comment -- I think you added this method here, so it should be better now, right?

[Frightera]

@ColCarroll Yes, it should be better. The last stable version of TFP was 0.15 when this method was implemented. I wanted to point out that it is still an issue with TFP 0.16.

Also, issue #1505 can be closed now.

Thanks.