This reverts commit 1814576c.

- pycuda initialization was also affected by the bug from CUDA 7.0

- fixing an issue introduced with CUDA-TK 7.0, when automatic selection
  of free gpu by os / library no longer works. Library always selects '0',
  which leads to crash in case of 'exclusive mode'.
- this fix is inspired by Dan Povey's fix for Kaldi:
  https://github.com/kaldi-asr/kaldi/commit/6548565445167e00125848f91d7da5f3f949b2a2
- it does a loop over gpus until a free gpu is taken.

Implement GpuAdvancedSubtensor1 for gpuarray

elemwise opt

Scan inplace opt

Use shape_of instead of explicitly inserting Shape_i

d3viz: Interactive visualization of Theano compute graphs

Add a 'name' argument to '.copy' method of Theano variables

These Shape_i nodes will not always be optimized away otherwise, and
they prevent local_shape_to_shape_i from being applied.
Instead, we manually apply the equivalent of local_shape_to_shape_i,
by using shape_of[...] directly.

Whitespace needed in theano_vs_c.txt

Fixed the docstring for the scan_opt.

Fix a non-ascii character in a comment in test_flake8.py.

Add some minimal tests for cudnn softmax.

Enable merging of GpuDnnPoolDesc

fix gh-3377 bug introduced by 41a8e89b merged by 7320e1b1. GpuElem…

fix gh-3377 bug introduced by 41a8e89b merged by 7320e1b1. GpuElemwise with non contiguous inputs give bad results

Change default value of reoptimize_unpickled_function

Remove the assert in PyGpuArray_DEV_DATA since it might cost a bit too much to verify that condition everytime.