Type issue in gkmx()

[ChenhanYu]

/frame/gkmx.hpp

Notice that currently gkmx<...,TA,TB,TC,TV> has the following issues.

1. TV cannot be different from TC.
2. Type rules in macro kernels may be wrong.

According to the definition, gkmx only need to pass in an m-by-n of C in type TC, but when k > KC the temporary rank-KC update must be stored as an m-by-NC matrix as type TV. It it very unpleasant to allocate this temporary buffer, but currently I have not find a way to resolve this. GKRM will have the same issue later.

Maybe we can increase KC such that k will never be larger than KC when TC != TV detected.

Notice that gkmx_gpu.hpp does not have this problem. GEMM algorithm on GPU does not store rank-KC update back to the global memory. L1 cache on GPU can be manually controlled; thus, storing back in unnecessary.