[utkinis]

mapreduce over several arrays (AK.mapreduce(f, op, a, b, c)) reduces a Broadcasted object. The threaded host reduction can't take views of a Broadcasted, so it ran these reductions on a single thread, whatever -t was.

This indexes the Broadcasted as a LinearBroadcast (from #185) instead. That's a vector, so it gets chunked across tasks like any other array. Each argument is indexed with its own linear index, so no Cartesian index is decoded per element. Indexing a 2-D Broadcasted with an Int is about 10× slower per element. A broadcast too long for LinearBroadcast's 32-bit index (over 2^32 elements) still runs serially.

Timings, before (a serial mapreduce over the Broadcasted) and after, measured in the same process. Ryzen 9 9950X, Julia 1.12, KA 0.10.3; minimum of 30 runs, on a shared machine:

	threads	before	after
x * (y - z), 3 × 2^23 Float64	1	6.8 ms	5.1 ms
	4	6.6 ms	4.2 ms
	16	6.6 ms	4.0 ms
x * y, 2 × 2048×4096 Float64	1	8.4 ms	3.5 ms
	4	8.4 ms	2.7 ms
	16	8.3 ms	2.6 ms
sin(x) * cos(y), 2 × 2^23 Float64	1	71 ms	73 ms
	4	68 ms	18 ms
	16	70 ms	6.7 ms

The first two cases are limited by memory bandwidth from about two threads up; the third shows the scaling.

The PR also adds a test that runs a 3-D multi-source reduction across several tasks, and a multi-source case to benchmark/mapreduce_1d.jl.