DArray: Add accumulate, cumsum and cumprod

[aalquwayfili]

Fixes #770.

accumulate had no DArray method, so Base walked it element by element. This adds one, and cumsum/cumprod use it.

Each partition is scanned on its own, then the last slice of the partitions before it is carried in along dims. The carries are one slice each, so only they run in order and the scans stay parallel. Like reduce this needs an associative op, so accumulate(-, DA) won't match Base.

One catch: only the first partition gets init, so with Int8 data and init=0 the later partitions overflow if you do nothing about it:

v = fill(Int8(100), 10)
collect(accumulate(+, distribute(v, Blocks(3)); init=0))
# without widening: [100, 200, 300, 400, 244, 344, 444, 288, 388, 488]

So when init widens the eltype, the partitions get widened before the scan. There's a test for that now.

array/mapreduce passes on 1.10, 1.11 and 1.12 (484 tests), and the gpu suite on a Quadro RTX 8000 (248). With the change reverted the new tests fail on both.

Rough timing on 8 threads (the machine was busy): cumsum(DA; dims=1) on 8192x8192 takes about 61 ms, vs 305 ms for Base on the same Array.

cumsum!/accumulate! into an existing DArray still fail, I left those for later. Also added the three to the supported list in the docs.

[aalquwayfili]

Pushed a fix for the OpenCL and Metal jobs.

OpenCL.jl has no accumulate for CLArray, so even a plain accumulate(+, A::CLArray) hits scalar indexing, and DArray accumulate can't work there yet. I marked it @test_broken so it shows up if that changes. Metal has no Float64, so I skipped the init=0.0 test there.

The other red jobs look unrelated, most of them fail on master too.

[jpsamaroo]

Thanks so much!