Flatten buoyancy_gradients call

[charleskawczynski]

A step towards #2530.

[charleskawczynski]

I looked at the nsight report in https://buildkite.com/clima/climaatmos-target-gpu-simulations/builds/272, and buoyancy_gradients + the 3 hoisted ᶜgradᵥ-ᶠinterp calls are now 3.9 ms, so this PR cuts the time in half (xref: #2950 (comment)). The buoyancy gradients itself is only 547 μs, the hoisted ᶜgradᵥ-ᶠinterp calls are 1.5 ms, 177 μs, and 1.6 ms respectively. So, I think we can open an issue in ClimaCore about the following patterns:

Maybe bad:

@. ᶜgradᵥ = ᶜgradᵥ(ᶠinterp(get_single_scalar(big_struct)))

Maybe bad:

@. ᶜgradᵥ = ᶜgradᵥ(ᶠinterp(foo(something)))

Maybe catastrophic:

@. single_scalar = foo(ᶜgradᵥ(ᶠinterp(get_single_scalar(big_struct))), ᶜgradᵥ(ᶠinterp(get_single_scalar(big_struct))))

Maybe catastrophic:

@. single_scalar = foo(ᶜgradᵥ(ᶠinterp(foo(something))), ᶜgradᵥ(ᶠinterp(foo(something))))

Good:

@. ᶜgradᵥ = ᶜgradᵥ(ᶠinterp(cfield))

We can make some reproducers for this in ClimaCore and iterate more quickly on improving things.

What I suspect was the culprit for this slowdown, is both/either not using shared memory, or register spills, due to complexity. Not using shared memory could have even resulted in register spills.

We should probably make our FD use shared memory, but this is at least an improvement, so I'll merge.

[charleskawczynski]

Also, just to note, the first commit didn't seem to have much of an impact on performance-- it was the hoisting that made the improvement.