Added CUB memory pool. CNMEM pool kept as default if USE_CNMEM set.

[borisfom]

Checking in as single squashed commit - trying to extract individual commits for CNMEM unintegration and CUB introduction appeared to be impractical.

[lukeyeager]

The Make build seems to work fine. The build succeeds and all tests pass whether you build with CNMeM or not.

But with CMake I'm getting this error whether I tell it to use CNMeM or not:

$ mkdir build && cd build
$ cmake ..
$ make -j12
...
/home/lyeager/caffe/nv-0.14.0-rc.1/src/caffe/util/gpu_memory.cpp:14:38: fatal error: cub/cub/util_allocator.cuh: No such file or directory
 #include "cub/cub/util_allocator.cuh"
                                      ^
compilation terminated.
make[2]: *** [src/caffe/CMakeFiles/caffe.dir/util/gpu_memory.cpp.o] Error 1
make[2]: *** Waiting for unfinished jobs....
make[1]: *** [src/caffe/CMakeFiles/caffe.dir/all] Error 2
make: *** [all] Error 2

[lukeyeager]

When I use Make build this, I can't tell that it's ever actually using a memory allocator - either CNMeM or CUB. I don't see it in the ldd output of the executable, and when I try to train a model the memory utilization is low - about 14% instead of 95%.

[lukeyeager]

Now that #69 is merged, I can see that CNMeM now utilizes 80% of memory instead of 95%. Is that expected?