Minimal implementation of new GEMM kernel

Getting this reviewed before adding customizations for NNC (batching, fused bias).