System information

• TensorFlow version (you are using): 1.3+ or 2.0a
• Are you willing to contribute it (Yes/No): Yes

Describe the feature and the current behavior/state.
Currently there are two "ways" to normalize vectors, matrices, and tensors.

1. tf.math.l2_normalize
2. tf.norm

The first one directly normalizes the input tensor and returns it.
The latter only computes the norm and the user has to do the division to obtain a normalized tensor.

I propose to add a function called tf.math.normalize which has the same/similar args as tf.norm. It computes the normalized tensor as tf.math.l2_normalize does but supports all the normalizations that tf.norm supports.

Will this change the current api? How?
No, it just adds a cleaner interface to do normalization.
In the long run tf.math.l2_normalize would be obsolete, though. It's implementation could be removed and it could use tf.normalize internally.

Who will benefit with this feature?
When looking for a way to perform l1 normalization I only found l2_normalize and tf.norm. I then expected there to exist a l1_normalize function as well which is not the case.
From my point of view this is an inconsistency that people will encounter from time to time.

I would propose to only add this to tensorflow 2.0 and maybe also remove tf.math.l2_normalize in that case.