map lacks specializations for sparse matrices/vectors

[Sacha0]

map lacks specializations for sparse matrices/vectors (in contrast to broadcast):

julia> VERSION
v"0.6.0-dev.1234"

julia> @which map(sin, speye(4))
map(f, A::Union{AbstractArray,AbstractSet,Associative}) in Base at abstractarray.jl:1784

julia> @which map(sin, spzeros(4))
map(f, A::Union{AbstractArray,AbstractSet,Associative}) in Base at abstractarray.jl:1784

julia> @which map!(sin, speye(4))
map!{F}(f::F, A::AbstractArray) in Base at abstractarray.jl:1775

julia> @which map!(sin, spzeros(4))
map!{F}(f::F, A::AbstractArray) in Base at abstractarray.jl:1775

This lack results in poor performance (see jump-dev/JuMP.jl#889). Ref. discussion in #7010. Best!

[stevengj]

It would be easy to add map(f, args::AbstractSparseVector...) = broadcast(f, args..) and map(f, args::AbstractSparseMatrix) = broadcast(f, args), since I think the semantics of map and broadcast agree for multiple sparse-vector arguments or a single sparse-matrix argument.

[simonster]

The semantics do not agree if some but not all sparse vectors have only a single element. But it seems like we could handle this by first checking dimensions and then calling broadcast if they match.

[tkelman]

Do any existing map methods assume purity?

[Sacha0]

I would probably go the other direction --- implement special cases of broadcast (specifically those where dimensions agree) via map, or implement both broadcast and map via common underlying methods. Best!

[Sacha0]

#19438 provided map/map! specializations for SparseMatrixCSCs (less a map!(f, A::SparseMatrixCSC) method), partially fixing this issue.

[tkelman]

assuming the rest was closed by #19690