Update. The GPU code does not more use `GpuArray_move` as this function seems to be too slow when copied data are not contiguous.

Now basic views are created and a view_map is declared for the op (on GPU only).

To come up:
* re-read the code to find some optimizations.
* check if using number of splits as a params op can be useful.

Method: for each split, a view is created and then copied into an output.

To come up:
* Add C code for GPU.
* Check if `self.len_splits` can be defined as a param op.

GpuArray_take1 requires contiguous inputs.

Import numpy gpuarray

Have "theano-cache list" include compilation times

Abstract Ops for batch normalization

(small fix) Update `theano-cache` to prevent an error message printed when `device=cuda` on Windows.

flag `device` is set to `cuda` on Windows.

Implement indexing with empty list

Implement conv2d_transpose convenience function

Add exception when epsilon is too small.

Use GraphToGPU for the new BN opt. Also remove the LocalGroupDB, as it add extra overhead and wasn't useul as there is only 1 GPU version.