Refactoring of gemm, adding faster kernel

This change gets rid of all non-batch functors, modularizes
duplicated code, and implement non-batches functions as calls
to batched functors with trivial constexpr batch indexer.

This change also adds faster gemm kernel that threads of N,M space,
and accumulates entire range of K in single work-item.

Dispatch logic changed too, we dispatch to thead-K kernel only if
(n,m) space is sufficiently small.

Author: oleksandr-pavlyk