removed repetition from local tile sum knl because of weird performance (compiler may be doing some kind of loop optimization

created new dg (naive vs. tiled) demo that utilizes training kernels specifically made to match dg access patterns

created matmul naive vs tiled demo v2, which uses a set of kernel generators specifically designed to match the matmul access patterns

in matmul naive vs tiled demo, now using training kernels that produce access-to-footprint-ratios that more closely match test kernels