Implement the grad for inner dimensions in MaxAndArgMax.
The pull request is fine. 

There is one question about gradient of max, that I'm not sure how to tackle, but the changes introduced in this code do not change the behaviour of Theano in this respect. 

The problem is how should the grad of max look if there are several instances of max ? Should all get the full gradient (like it is happening now), should only one of them receive the gradinet ?