Implement greedy scheduler

[MasonProtter]

Implements #11

Like split=:scatter, this option is a bit scary to expose for tmapreduce because it also requires that the reducing op be commutative.

I almost wonder if we should do some sort of argument validation to detect if the user gave us a non-commutative reducing operator...

e.g.

julia> treduce(vcat, [:a, :b, :c, :d, :e, :f, :g])
7-element Vector{Symbol}:
 :a
 :b
 :c
 :d
 :e
 :f
 :g

julia> treduce(vcat, [:a, :b, :c, :d, :e, :f, :g], schedule=:greedy, init=Symbol[])
7-element Vector{Symbol}:
 :b
 :f
 :a
 :g
 :d
 :e
 :c

This is not a problem for tforeach, tmap(f, OutputElementType, A), or tmap!, but it is a concern for regular tmap, as well as tmapreduce, treducemap, and treduce.

One other potential problem is that this algorithm is prone to giving errors if you run it on a small array and don't provide an init argument:

julia> treduce(vcat, [:a, :b, :c, :d, :e, :f, :g], schedule=:greedy)
ERROR: TaskFailedException
Stacktrace:
[...]
    nested task error: MethodError: reducing over an empty collection is not allowed; consider supplying `init` to the reducer
    Stacktrace:
[...]

this is because we're doing

    ch = Channel{Tuple{eltype.(Arrs)...}}(0; spawn=true) do ch
        for args ∈ zip(Arrs...)
            put!(ch, args)
        end
    end
    tasks = map(1:ntasks) do _
        @spawn mapreduce(op, ch; mapreduce_kwargs...) do args
            f(args...)
        end
    end

so if one task gets spawned after the channel gets emptied, then we're in trouble.

[carstenbauer]

Like split=:scatter, this option is a bit scary to expose for tmapreduce because it also requires that the reducing op be commutative.

I almost wonder if we should do some sort of argument validation to detect if the user gave us a non-commutative reducing operator...

I think documenting this requirement clearly for these options should be fine. After all, you have to actively opt for them. As for testing the commutative property, I would think that that's hard (perhaps even impossible) to do in general and also can be costly.

[MasonProtter]

Do you think the docstrings I put in are sufficiently clear?

[carstenbauer]

I think so, yes. But I'll revisit them when I set up Documenter (today).