Explicitly specify the output ndim in reshape

Add in the doc the optional dependencies of pycuda and skcuda.

Add test for the previous fix and add more check to prevent other crash with bad neib_shape or neib_step input value.

Fix dnn conv grad issues

The whole code expects the shape to be of length 4, and the output to be
4D already

Updates to solve potential problems with compiler-specific casting

Print scan profile again

Remove the old benchmark directory.

Remove the old benchmark directory. They use the module system that was removed years ago. Now the benchmark are outside Theano.

Fix various GPU unit tests on Windows

Modify error printing to make it more explicit when calling `GpuArray_maxandargmax()`.

Fix doc version switch and unverbose PR GPU tests

 Fix NaNGuardMode to permit None variables

rather than fct_callcount, which stays at 0

Fix GPU test_pool on Windows.

Allow for pooling of empty batch (like for conv in #3715).

theano/gpuarray/tests/test_subtensor.py:G_subtensor.test_noncontiguous_idx

The old big size was used to check if reduction is really faster
on GPU than on CPU with big data, but it made NVIDIA driver crash
on Windows, and I think it is not really useful for simple unit
testing.

[CRASH] Fix local_gpua_advanced_incsubtensor to return variable on the GPU

Cast ws, strides and pads to 64 bits integers into both Python and C code
for all these ops.

Make sure that ws, strides and pads for GpuPool, GpuMaxPoolGrad and GpuAveragePoolGrad are `npy_intp` arrays.

Fix local_gpua_advanced_incsubtensor to return variable on the GPU. op_lifer expect that. fix gh-5589

Prevent pkl_utils from stomping higher sys recursion limits