Fix memory leak in scan related to the input and output storage of the inner function.

Fix blocksparse optimization that could give bad results in some cases

[ENH] Add SliceConstant, better stack trace, fast_compile_gpu

Also add patch for numpy API fixes to help people playing with scan_perform.pyx

This is because if we exclude the 'gpu' tag, we don't want those opt to also be enabled.

Remove internal Theano stack level from variable creating. This make them more readable to the user.

The added test currently fail.

[MRG, ENH]Some crash/warning fix and skip some tests when g++ isn't available.

fix gh-2138

Fix destructor handling in CDataType for python 3/Capsules.

Include and library flags in nvcc.flags in config not passed to nvcc correctly

Fix crash when numpy.distutils.__config__ is present but None

Added -I and -L to the list of argument patterns from nvcc.flags that are not presented to nvcc in comma-separated form.

Implemented cuDNN softmax but it is not currently passing tests.

Incremeneted GpuDnnSoftmax version and retested as an extra sanity test of my recent refactoring.

[MRG, BUG] Scan Opt fixes, enhencements