Fix sum(bc::Broadcasted; dims = 1, init = 0) (#43736)

This PR make `has_fast_linear_indexing` rely on `IndexStyle`/`ndims` to
fix `mapreduce` for `Broadcasted` with `dim > 1`.
Before:
```julia
julia> a = randn(100,100);

julia> bc = Broadcast.instantiate(Base.broadcasted(+,a,a));

julia> sum(bc,dims = 1,init = 0.0) == sum(collect(bc), dims = 1)
ERROR: MethodError: no method matching LinearIndices(::Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{2}, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}, typeof(+), Tuple{Matrix{Float64}, Matrix{Float64}}})
```
After:
```julia
julia> sum(bc,dims = 1,init = 0.0) == sum(collect(bc), dims = 1)
true
```

This should extend the optimized fallback to more `AbstractArray`. (e.g.
`SubArray`)

Test added.

base/reducedim.jl

@@ -196,11 +196,8 @@ end
 
 ## generic (map)reduction
 
-has_fast_linear_indexing(a::AbstractArrayOrBroadcasted) = false
-has_fast_linear_indexing(a::Array) = true
-has_fast_linear_indexing(::Union{Number,Ref,AbstractChar}) = true  # 0d objects, for Broadcasted
-has_fast_linear_indexing(bc::Broadcast.Broadcasted) =
-    all(has_fast_linear_indexing, bc.args)
+has_fast_linear_indexing(a::AbstractArrayOrBroadcasted) = IndexStyle(a) === IndexLinear()
+has_fast_linear_indexing(a::AbstractVector) = true
 
 function check_reducedims(R, A)
     # Check whether R has compatible dimensions w.r.t. A for reduction

test/broadcast.jl

@@ -979,6 +979,10 @@ end
     @test sum(bc, dims=1, init=0) == [5]
     bc = Broadcast.instantiate(Broadcast.broadcasted(*, ['a','b'], 'c'))
     @test prod(bc, dims=1, init="") == ["acbc"]
+
+    a = rand(-10:10,32,4); b = rand(-10:10,32,4)
+    bc = Broadcast.instantiate(Broadcast.broadcasted(+,a,b))
+    @test sum(bc; dims = 1, init = 0.0) == sum(collect(bc); dims = 1, init = 0.0)
 end
 
 # treat Pair as scalar: