Softmax operator shape inference

[roywei]

after apache/mxnet#14661 and apache/mxnet#14869.
Softmax operator has more strict partial shape inference. This will result in the ndim() of softmax input maybe unknown during dimension check. It will be inferred at a later time.

I have created PR for the case of ndim == 0 in for mxnet in #232.
I'm keeping this issue open to see if we can find alternatives of modifying Keras front end code.