Add CNMeM support

[slayton58]

Add support for the CNMeM memory pool by abstracting out cudaMalloc and cudaFree operations into calls to the singleton (Caffe::mallocGPU and Caffe::freeGPU). Memory pools are lazily initialized on first allocation.

Notes:

• Currently takes 90% of available GPU memory at memory pool initialization - this value can be tweaked
• Serializes backwards convolution calls to simplify memory management - instead of running all groups in parallel and bgrad, dgrad and wgrad in parallel within each group (all using streams), now uses a single stream. This allows all remaining memory in the pool to be used for any given convolution instead of sharing. This is easy to change back if desired.
• Enable using Caffe::set_memory_pool(true, gpus); call (gpus is of type std::vector), currently only enabled in the main caffe executable
• Currently tests do not use the memory pool (they could be added for tests that use GPUDeviceTest)

[lukeyeager]

Added to the caffe-0.13 branch in version v0.13.1
https://github.com/NVIDIA/caffe/commits/v0.13.1