fix gh-4036. We need the shape and we don't compute it. It wasn't timmed, so don't try this optimization.

Cudnn version, print it, warn when too new, make mandatory in the new back-end.

Fix segfault for very big size. I did some timing and there is no di…

Fix segfault for very big size. I did some timming and there is no difference that is higher then noise. fix gh-4020

Replace deleted pages with link to new pages

Added the commit instruction in dev_guide submit text

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".