Update the versions tested on travis

Single stream flag

Port FFTs to gpuarray backend

Verify grad speedup

Ccw 4483 indent fix

ones_like and zeros_like dtype parameter is documented

Useless alloc

Don't use DebugMode for the numerical gradient in verify_grad.

Exclude optimization to keep old test working

Set cuda/test_dnn to run conv references in fast run

Unbreak travis.