Added new ScanPrefix accumulate algorithm

[anicusan]

Added second accumulate algorithm using coupled lookback of pre-scanned prefixes (=> one extra kernel launch), with that ScanPrefixes algorithm becoming the default on Metal.

This fixes the decoupled-lookback issue on Metal.

[anicusan]

We're still faster than the current default Metal accumulate:

using BenchmarkTools
using Metal
import AcceleratedKernels as AK

using Random
Random.seed!(0)

function akacc(v)
    va = AK.accumulate(+, v, init=zero(eltype(v)), block_size=1024)
    Metal.synchronize()
    va
end

function baseacc(v)
    va = accumulate(+, v, init=zero(eltype(v)))
    Metal.synchronize()
    va
end

v = MtlArray(rand(1:100, 10_000_000))

# Correctness checks
va = akacc(v) |> Array
vb = baseacc(v) |> Array
@assert va == vb

# Benchmarks
println("Base vs AK")
display(@benchmark baseacc($v))
display(@benchmark akacc($v))

And timings:

julia> include("accumulate_benchmark.jl")
Base vs AK
BenchmarkTools.Trial: 603 samples with 1 evaluation.
 Range (min … max):  3.369 ms … 52.091 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     7.746 ms              ┊ GC (median):    0.00%
 Time  (mean ± σ):   8.300 ms ±  6.782 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

  █▆▄▁  ▁ ▅▇▅▂                                                
  █████████████▇▅▇▇▆▇▅▆▅▄▅▄▄▁▄▁▁▅▁▅▁▆▆▄▆▄▇▅▄▄▄▄▁▁▁▁▁▁▁▁▁▄▄▅▆ ▇
  3.37 ms      Histogram: log(frequency) by time     36.9 ms <

 Memory estimate: 45.41 KiB, allocs estimate: 1568.
BenchmarkTools.Trial: 644 samples with 1 evaluation.
 Range (min … max):  4.535 ms … 35.595 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     5.928 ms              ┊ GC (median):    0.00%
 Time  (mean ± σ):   7.770 ms ±  4.089 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

  █▅▁    ▁▂▁▂▄▄▃▂                                             
  ███▇▆▁▆██████████▇▄▆▆▇▆▇▅▇▅▆▁▁▅▁▁▁▄▄▁▁▄▁▁▄▄▁▁▁▅▁▁▁▁▁▁▄▄▁▁▄ ▇
  4.53 ms      Histogram: log(frequency) by time     27.6 ms <

 Memory estimate: 16.63 KiB, allocs estimate: 565.