Add methods to bypass ConditionalOperation when condition==nothing

[tomchor]

Related to #4633

Wrote this short script to time the slowdown we have (compared to directly accessing interior(u) when using mean(u) and Integral(u):

using Oceananigans, BenchmarkTools

grid = RectilinearGrid(size=(8, 8, 8), extent=(1, 1, 1), halo=(1, 1, 1));
u = XFaceField(grid);

bmean_baseline = @benchmark mean(interior(u))
bmean = @benchmark mean(u)
slowdown_mean = minimum(bmean.times) / minimum(bmean_baseline.times)
@info "Slowdown using mean(u): $slowdown_mean"

ū = Field(Average(u))
bavg = @benchmark compute!(ū)
bsum_baseline = @benchmark sum(interior(u))
slowdown_avg = minimum(bavg.times) / minimum(bsum_baseline.times)
@info "Slowdown using compute!(ū): $slowdown_avg"

On main this produces

[ Info: Slowdown using mean(u): 8.200285408490902
[ Info: Slowdown using compute!(ū): 6.156982321501987

On this branch

[ Info: Slowdown using mean(u): 5.059299971073185
[ Info: Slowdown using compute!(ū): 1.8239949351060463

So this definitely speeds things up (when ConditionalOperation isn't needed), although there is more work to be done.

[glwagner]

Why does the slow down occur?

[glwagner]

If necessary we should do this, but I wonder if a better design would simply avoid creating the pointless ConditionalOperation in the first place?

[glwagner]

Something is sketchy because the conditional operation should have no effect due to the constant here:

Oceananigans.jl/src/AbstractOperations/conditional_operations.jl

Line 156 in b400322

@inline evaluate_condition(::Nothing, i, j, k, grid, args...) = true

[glwagner]

and moreover for example:

Oceananigans.jl/src/AbstractOperations/conditional_operations.jl

Line 146 in b400322

@propagate_inbounds Base.getindex(co::NoFuncNoConditionCO, i, j, k) = getindex(co.operand, i, j, k)

although I am not sure that method is being hit properly, maybe it isn't.

[simone-silvestri]

I think I found one issue. mean(interior(u)) is slower than it should be. Apparently, all the maps we have in interior are slowing down considerably all the computations, I ll open a different PR, then we can revisit this one. After fixing the maps, the additional cost of adding a conditional operation is concentrated in the Field constructor.

[simone-silvestri]

See #4668

[tomchor]

Closing this since #4795 made this PR obsolete