Faster sum(BroadcastArray)

[mcabbott]

This simply makes the complete sum of a BroadcastArray call sum(A,dims=1:N) as this seems to be faster. For example, without this PR:

julia> using LazyArrays, BenchmarkTools

julia> A = rand(100,100); B = rand(100); C = rand(100);

julia> @btime sum($A .* $B .* $C')
  5.316 μs (2 allocations: 78.20 KiB)
1220.0468263958778

julia> @btime sum(BroadcastArray(*, $A, $B, $C'))
  26.259 μs (11 allocations: 256 bytes)
1220.046826395877

julia> @btime sum(BroadcastArray(*, $A, $B, $C'), dims=(1,2))
  3.062 μs (12 allocations: 416 bytes)
1×1 Array{Float64,2}:
 1220.0468263958776

Another example in #16. That issue also has a puzzle about B' vs reshape(B,1,:) which this does nothing about.

[codecov-io]

Codecov Report

Merging #18 into master will increase coverage by 0.06%.
The diff coverage is 100%.

@@            Coverage Diff            @@
##           master     #18      +/-   ##
=========================================
+ Coverage   49.74%   49.8%   +0.06%     
=========================================
  Files          11      11              
  Lines         782     783       +1     
=========================================
+ Hits          389     390       +1     
  Misses        393     393

Impacted Files	Coverage Δ
src/lazybroadcasting.jl	45.83% <100%> (+2.35%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 4559ba2...5c99cbb. Read the comment docs.