Allow any f for sum/minimum/maximum(f, v::AbstractSparseVector)

[alyst]

At the moment only f=abs/abs2 have specialized implementations of sum/minimum/maximum(f, v::AbstractSparseVector), the other fs are handled by the generic AbstractVector methods.
This PR extends sparse-specific methods to any f.

The commit was extracted from #28535 since it should be a non-breaking backportable change.

cc @mbauman

[simonbyrne]

This assumes f(zero(eltype(x))) is pure, which may not be the case. As far as I know, we don't have a way to check if a particular function or method is pure, but we could add a trait for this?

[StefanKarpinski]

I think that's a fairly reasonable assumption though. What kinds of types are you thinking about where that would fail? BigInt and BigFloat because they allocate?

[mbauman]

Yeah, we make that assumption throughout the sparse library. It's not the zero part that's the problem, it's the user-passed f — which could depend on global state (e.g. f(x) = x + rand(0:1)).

[mbauman]

I know it's unrelated, but maybe throw an @test_broken sum(t -> true, zeros(Float64, 0)) === 0 right next to this line so that way when one gets fixed the author will find the other.

[simonbyrne]

Yeah, we make that assumption throughout the sparse library. It's not the zero part that's the problem, it's the user-passed f — which could depend on global state (e.g. f(x) = x + rand(0:1)).

So we do...

julia> map(x -> rand(), sparse(zeros(5)))
5-element SparseVector{Float64,Int64} with 5 stored entries:
  [1]  =  0.658879
  [2]  =  0.658879
  [3]  =  0.658879
  [4]  =  0.658879
  [5]  =  0.658879

[simonbyrne]

Could this be done more generally by hooking into the mapreduce machinery?

[ViralBShah]

@alyst This is an old PR. Any chance you'd be able to update it? It would be great to get this in, and then get the mapreduce stuff later on.

[alyst]

tester_win32 failure is unrelated (permission denied), while the SparseVector tests had passed.

[vtjnash]

@ViralBShah is this good to go?