Add mapreduce.

[maleadt]

Initial implementation. Performs a little better than CUDA.jl already, at least on the simple cases I tested:

julia> A = CUDA.rand(1024, 1024, 1024);


julia> CUDA.@profile trace=true sum(A)

Device-side activity: GPU was busy for 7.9 ms (33.76% of the trace)
┌─────────┬─────────┬────────┬──────┬──────────────────┬───────────────────────┐
│    Time │ Threads │ Blocks │ Regs │       Shared Mem │ Name                  │
├─────────┼─────────┼────────┼──────┼──────────────────┼───────────────────────┤
│  7.9 ms │    1024 │     84 │   60 │ 128 bytes static │ simt_reduce_kernel    │
│ 2.38 µs │      96 │      1 │   70 │ 128 bytes static │ simt_reduce_kernel    │
└─────────┴─────────┴────────┴──────┴──────────────────┴───────────────────────┘


julia> CUDA.@profile trace=true sum(ct.Tiled(A))

Device-side activity: GPU was busy for 5.22 ms (2.55% of the trace)
┌───────────┬─────────┬─────────┬──────┬───────────────────┬─────────────────────┐
│      Time │ Threads │  Blocks │ Regs │        Shared Mem │ Name                │
├───────────┼─────────┼─────────┼──────┼───────────────────┼─────────────────────┤
│   5.22 ms │     128 │ 1×1×128 │   76 │ 32.035 KiB static │ tiled_reduce_kernel │
│   1.67 µs │     128 │       1 │   21 │   28 bytes static │ tiled_reduce_kernel │
└───────────┴─────────┴─────────┴──────┴───────────────────┴─────────────────────┘

It's a bit unfortunate there's so much duplication with GPUArrays, but making Tiled <: AbstractGPUArray probably breaks more than it is worth.

Implementation could be much improved, e.g., by using atomic reductions instead of a multi-phased reduction.