Update documentation of the tf.nn.softmax function

[T-Almeida]

This PR addresses the issue #53112

As discussed, the tf.nn.softmax documentation states that:

This function performs the equivalent of

softmax = tf.exp(logits) / tf.reduce_sum(tf.exp(logits), axis)

However the above computation is wrong for tensors with rank greater than 1.

I belive that the correct description would be:

softmax = tf.exp(logits) / tf.reduce_sum(tf.exp(logits), axis, keepdims=True)

[T-Almeida]

Hi, I noticed that the line is longer than 80 characters, any suggestion on how to rewrite it in a two-line format?

P.S. I also tried to run tf_doctest.py (as indicated here: https://www.tensorflow.org/community/contribute/docs_ref) but it throws "AttributeError: module 'tensorflow.python.ops.logging_ops' has no attribute 'enable_interactive_logging"

[gbaned]

@T-Almeida Can you please address PyLint errors? Thanks!

[T-Almeida]

Hi @gbaned, I made two changes in order to fix the PyLint erros, see if its okey