Add missing optimization for iterate(::LogicalIndex{<:CartesianIndex,<:BitArray})

[N5N3]

This PR aims to make view(A, A .> 0) faster when IndexStyle(A) isa IndexCartesian.
Some benchmark:

a = randn(128, 32);
A = view(a, axes(a)...);
A2 = view(reshape(a,32,32,4),1:32,1:32,1:4);
p = a .> 0;

@btime view($a, $p); # 1.780 μs (4 allocations: 16.56 KiB)
@btime view($A, $p); # 16.700 μs (2 allocations: 32.92 KiB)
@btime view($A2, reshape($p, size($A2))); # 13.500 μs (4 allocations: 49.47 KiB)

# After This PR
Base.collect(L::Base.LogicalIndex{Int,<:BitArray}) = findall(vec(L.mask))
Base.collect(L::Base.LogicalIndex{CartesianIndex{N},BitArray{N}}) where {N} = N >= 2 ? findall(L.mask) : [i for i in L]

@btime view($a, $p); # 1.710 μs (6 allocations: 16.66 KiB)
@btime view($A, $p); # 2.522 μs (2 allocations: 32.92 KiB)
@btime view($A2, reshape($p, size($A2))); # 3.638 μs (4 allocations: 49.47 KiB)

[N5N3]

Profile shows that the slow speed for IndexCartesian comes from the unoptimized fallback.
So I add the missing optimization with findall as a reference.
This will also make A[A.>0] = B faster.

[N5N3]

Update the Benchmark

using Random, BenchmarkTools
Random.seed!(0)
a = randn(128, 32);
A = view(a, axes(a)...);
A2 = view(reshape(a,32,32,4),1:32,1:32,1:4);
p = a .> 0;
@btime view($a, $p); # 1.630 μs  -->  1.360 μs
@btime view($A, $p); # 16.000 μs --> 2.500 μs
@btime view($A2, reshape($p, size($A2))); # 13.100 μs --> 3.638 μs

[ViralBShah]

Is this good to merge?

[N5N3]

We already have

julia/test/abstractarray.jl

Lines 458 to 462 in f090992

	mask = bitrand(shape)
	@testset "test logical indexing" begin
	@test B[mask] == A[mask] == B[findall(mask)] == A[findall(mask)] == LinearIndices(mask)[findall(mask)]
	@test B[vec(mask)] == A[vec(mask)] == LinearIndices(mask)[findall(mask)]
	mask1 = bitrand(size(A, 1))

So the added optimization should have been tested.