Global memory manager and preallocated buffer size

[ChenhanYu]

Notice that now we allocate packing buffer in a dynamic fashion inside our framework. The overhead seems ok but it may be good to preallocate the buffer with an initialized function such hmlp_init().

The reason that this is valuable is because there are some other temporary buffers we may need.

GSKS and GSKNN:
when k > KC, we need to store the rank-KC update.

GKMM:
when k > KC and TV != TC, we need to store the rank-KC update in type TV.

GKRM:
when k > KC and TV != TC, we need to store the rank-KC update in type TV.

GKRM on GPU:
we need an m-by-n/4 buffer in type TC to perform global reduction.

several references: (thanks Jianyu for providing the links)

http://stackoverflow.com/questions/29410064/aligned-dynamic-array-and-smart-pointer
http://www.codeproject.com/Articles/392576/Ideas-from-a-smart-pointer-Support-data-alignment