This concerns, max, argmax, min and argmin.
The case where ndim==1 was handled, but not ndim==0.

Don't print memory profile of theano function that use less then the theano flag ProfileMode.min_memory_size intermediate memory size.

make a test pass by accepting a higher (rtol=5e-6) diff threshold between cpu and gpu. This is expected as the rounding and order are not the same on cpu and gpu.

- fixes errors in conv_patch_stack_reduce when the entire kernel doesn't fit
  into shared memory

- lowers the shared memory requirement of conv_patch_stack_reduce