Speed up h_softmax when full output is requested.

add an assert statement in the test of the h_softmax to checks that fun_output_tg and fun_output are consistent

Add tp_repr to CudaNdarray. 

Move the extending_theano documentation out of tutorial

Temp disable warn64

Really skip the flake8 part in "everything but tensor".

Fix last DLT problem(s) with the new backend.

Add tp_repr to CudaNdarray. This make DebugMode error print the value instead of <CudaNdarray object at 0x7f229c2f6eb0>

A few small stuff

fixing make_node param passing for conv_grad (conv_transpose)

This was causing problems in the case of Pool, but I also fixed it elsewhere just to be sure.