Attempt to optimize `run_field_matrix_solver!` for CUDA

[petebachant]

This shows up as the top kernel in the "flagship" config (with the exception of 0M instead of 1M microphysics):

Above is from results/nsys/baseline.nsys-rep, commit a0bb502 of this repo.

I've tried a few modifications, which are summarized here. So far, none have sped up the simulation--only improved registers per thread and occupancy metrics.

[petebachant]

This presentation shared by @imreddyTeja may be relevant: https://www.nvidia.com/content/GTC/documents/1059_GTC09.pdf

[petebachant]

Strategy notes

• Use dycore_prognostic_EDMF_FieldMatrix to generate a realistic field matrix to use for benchmark case
• Adapt "FieldMatrixSolver Unit Tests" test to solve these matrices, optionally skipping if they're not tri-diagonal, since that appears to be the most common (drop anything that's not a mat3)
• Use helem=16 (or 30), npoly=3, velem=63 -- be careful about usage of test_spaces

[petebachant]

Tried implementing a parallel cyclic reduction solver and it was slower in the benchmarks with higher error, but ~3% faster in the flagship coupler simulation (commit; will need to implement permalinks to nsys reports, but they can be pulled with calkit).