reverted back the min to numpy.min

revert back to zip from izip.

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