Add regression test for nvcc.compiler_bindir (#5373)

Bug: gradient of ifelse with gpu inputs

Fix the buildbot.

Mixed fixes

gpuarray: Revert part of c6ffa460

Select the dnn convolution algorithm using actually available memory

crash fix. Add the missing inputs to the function call. It also need the real data. I don't sync as we don't care of the values.

Fix cudnn init in the new back-end. Remove duplicate place where we store cudnn version at the same time.

Also reserve 50M of device memory for other such handles.

Fix test that we don't move 4d tensor to the GPU

Gpu op for solve using cusolver (CCW #5191)

Fix issue #4476 related to BatchedDot.

Force install of the new commit of Theano

Otherwise, it is just using the cached one...

Use GpuAdvancedIncSubtensor1_dev20 more frequently in the new back-end