Improve RRTMGP.jl performance

[simonbyrne]

TODO: set target?

[sriharshakandala]

Targets for all-sky problem:
RRTMGP shortwave solve time on (1) V100 for 131,658 cols at 0.98 sec
RRTMGP longwave solve time on (1) V100 for 131,658 cols at 0.85 sec
These targets needs to be adjusted appropriately for P100 and A100 GPUs

[sriharshakandala]

Current status:
We are currently at 1.05 sec for the shortwave solver with 131,658 cols on a A100 GPU

[sriharshakandala]

Update OneScalar and TwoStream structs #434