sandbox/dan2: Modify CUDA matrix allocation so that it caches freed memory and...

sandbox/dan2: Modify CUDA matrix allocation so that it caches freed memory and returns cached answers if they are of the size required... this is to work around the extremee slowness of cudaMalloc and cudaMallocPitch on some platforms.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/dan2@3104 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8