Use less DataLayouts internals in DSS

[charleskawczynski]

This is a first step towards #2048.

[sriharshakandala]

Generally speaking, max_threads = 256 works well. If resource constraints (shmem size, register usage, etc.) permit, CUDA will automatically schedule multiple thread blocks on the same streaming multiprocessor and no additional intervention is required.
Using threads_via_occupancy does not necessarily guarantee optimality! Its up to the user to design the best threadblock configuration for their application, and I believe we should custom tune it, especially for performance critical kernels!

[charleskawczynski]

Generally speaking, max_threads = 256 works well. If resource constraints (shmem size, register usage, etc.) permit, CUDA will automatically schedule multiple thread blocks on the same streaming multiprocessor and no additional intervention is required. Using threads_via_occupancy does not necessarily guarantee optimality! Its up to the user to design the best threadblock configuration for their application, and I believe we should custom tune it, especially for performance critical kernels!

I'm fine with reverting that for now because it's not really part of the refactor. I am curious how this differs from the occupancy API