Inline optimization for destroyhandler.py

Fix bad change to gpuarray optimizations.

Add CNMeM in Theano to speed up CUDA allocation.

 - getroot
 - add_impact
 - get_impact

OpFromGraph.infer_shape()

This new version of CNMeM corresponds to the following commit
https://github.com/NVIDIA/cnmem/commit/2559e911ca5ad33c8a8aa7e5877345265115d963

It contains two bug fixes:
1/ One critical fix when the first device used by the library is
   not device 0. It could result in a call to cudaSetDevice(-1).
2/ One minor issue with a call to cnmemMalloc with a non-NULL
   pointer but a size of 0. It returned an CNMEM_STATUS_BAD_PARAM
   in the past. It is now fixed and it returns success.

I also changed the code in cuda_ndarray.cu to remove the extra
check when the size of the allocation was 0.

Update the code to support CNMeM (formerly known as Cumem).

flake8 for tensor/nnet/nnet.py