Add matrix-field benchmark

[charleskawczynski]

This PR adds a matrix field benchmark that is currently impressively slow on the A100 (on pretty low resolution):

BenchmarkTools.Trial: 107 samples with 1 evaluation per sample.
 Range (min … max):  43.442 ms … 48.436 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     47.463 ms              ┊ GC (median):    0.00%
 Time  (mean ± σ):   46.995 ms ±  1.086 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

                                                 ▄▄▇▃█         
  ▃▁▁▁▃▃▁▁▁▃▁▁▁▁▁▄▁▃▃▄▁▁▁▃▃▄▇▁▄▁▅▃▃▄▁▃▃▁▃▁▁▁▃▃▆▆▇█████▃▇▄▁▁▁▃ ▃
  43.4 ms         Histogram: frequency by time        48.4 ms <

 Memory estimate: 1.22 KiB, allocs estimate: 59.

For reference / comparison, pointwise operations for a single assign can be (easily) as low as 100μs. This benchmark only requires 3 reads (2 fields +1 boundary read) and 1 write (4 reads+writes total). So, I'd estimate that this should take at most 400μs (100x faster).

I'd like to do a couple things:

• make a utils file, so that we can use this example for running nsight compute, as well as a benchmark
• make the resolution flexible, so that we can see how this performs with different resolutions
• make this somehow reusable for our shared memory unit tests, since adding shmem support for these operations should (likely) improve performance.

This is related to CliMA/ClimaLand.jl#1097.