Add FieldVector microbenchmarks and improve fieldvector broadcast performance

[charleskawczynski]

Closes #2067. This turned out to be pretty easy: any fieldvector operation is embarrassingly parallel. We can leverage NonExtrudedBroadcasted, forward everything to the backing arrays, and just linearly index everywhere. This puts us in a pretty good shape for fieldvector operations, and it'll be even better at lower resolution since we parallelize across field variables.

GPU

main

N reads-writes: 5,  Float_type = Float64, Device_bandwidth_GBs=2039
┌─────────────┬───────────────────────────────────┬─────────┬─────────────┬──────────────┬────────┐
│ funcs       │ time per call                     │ bw %    │ achieved bw │ problem size │ n-reps │
├─────────────┼───────────────────────────────────┼─────────┼─────────────┼──────────────┼────────┤
│ FieldVector │ 1 millisecond, 104 microseconds   │ 54.161  │ 1104.34     │ (32745600,)  │ 4475   │
│ FieldVector │ 307 microseconds, 597 nanoseconds │ 48.6243 │ 991.45      │ (8186400,)   │ 10000  │
└─────────────┴───────────────────────────────────┴─────────┴─────────────┴──────────────┴────────┘

this PR

N reads-writes: 5,  Float_type = Float64, Device_bandwidth_GBs=2039
┌─────────────┬───────────────────────────────────┬─────────┬─────────────┬──────────────┬────────┐
│ funcs       │ time per call                     │ bw %    │ achieved bw │ problem size │ n-reps │
├─────────────┼───────────────────────────────────┼─────────┼─────────────┼──────────────┼────────┤
│ FieldVector │ 757 microseconds, 513 nanoseconds │ 78.9779 │ 1610.36     │ (32745600,)  │ 6491   │
│ FieldVector │ 202 microseconds, 849 nanoseconds │ 73.7335 │ 1503.43     │ (8186400,)   │ 10000  │
└─────────────┴───────────────────────────────────┴─────────┴─────────────┴──────────────┴────────┘

CPU

main

N reads-writes: 5,  Float_type = Float64,
┌─────────────┬──────────────────────────────────┬─────────────┬──────────────┬────────┐
│ funcs       │ time per call                    │ achieved bw │ problem size │ n-reps │
├─────────────┼──────────────────────────────────┼─────────────┼──────────────┼────────┤
│ FieldVector │ 32 milliseconds, 45 microseconds │ 38.0674     │ (32745600,)  │ 33     │
│ FieldVector │ 7 milliseconds, 792 microseconds │ 39.1372     │ (8186400,)   │ 529    │
└─────────────┴──────────────────────────────────┴─────────────┴──────────────┴────────┘

This PR

N reads-writes: 5,  Float_type = Float64,
┌─────────────┬───────────────────────────────────┬─────────────┬──────────────┬────────┐
│ funcs       │ time per call                     │ achieved bw │ problem size │ n-reps │
├─────────────┼───────────────────────────────────┼─────────────┼──────────────┼────────┤
│ FieldVector │ 23 milliseconds, 218 microseconds │ 52.5398     │ (32745600,)  │ 43     │
│ FieldVector │ 5 milliseconds, 752 microseconds  │ 53.0143     │ (8186400,)   │ 828    │
└─────────────┴───────────────────────────────────┴─────────────┴──────────────┴────────┘

[charleskawczynski]

Interesting, this shows that ClimaLand is using fieldvectors in some untested edge case. I'm going to find out the difference and add some unit tests that exercise this.

[charleskawczynski]

This looks done! 🎉

[charleskawczynski]

Just to comment on why this took a bit longer to iron out: the land model seems to rely on recursively defined fieldvectors, which we support copyto! for. Doing this correctly is a bit tricky because we need to dispatch to a custom (and device-specific) method to better optimize. All that said, I'm happy with the result.