Fix tail effect on GPUs

[dennisYatunin]

This PR extends max_resident_blocks to account for register pressure and shared memory requirements, fixing the "tail effect" problem that was introduced in #2522 for 64-level pointwise kernels (see this comment for details).

• Code follows the style guidelines OR N/A.
• Unit tests are included OR N/A.
• Code is exercised in an integration test OR N/A.
• Documentation has been added/updated OR N/A.

[imreddyTeja]

This does fix the tail effect, but it does not give the 50% speedup that ncu thought it would. It still says the average active warps increased by 1. It does give a small speed up. I'm a bit concerned about how this will impact ClimaLand though. I'd guess the extra dictionary usage might impact latency