Remove setdiff in kfolds

[alexiscltrn]

Rather than using setdiff to compute training indices (which can allocate large vectors on big datasets), compute them analytically from fold sizes and offsets. This reduces memory usage and improves performance.

kfolds with setdiff

julia> @benchmark kfolds(Int(1e5), 5) seconds = 600
BenchmarkTools.Trial: 10000 samples with 1 evaluation per sample.
 Range (min … max):  22.937 ms … 272.079 ms  ┊ GC (min … max): 0.00% … 80.58%
 Time  (median):     25.023 ms               ┊ GC (median):    3.32%
 Time  (mean ± σ):   26.898 ms ±   9.199 ms  ┊ GC (mean ± σ):  5.33% ±  5.44%

  ▃▄█▇▆▆▅▄▄▄▄▃▂▂▂▁▁▁▁                                          ▂
  ████████████████████████▇▇▇▇▆▆▆▅▆▆▆▄▆▆▅▅▅▆▄▆▆▅▁▅▅▅▅▁▅▄▅▃▃▄▅▄ █
  22.9 ms       Histogram: log(frequency) by time      57.2 ms <

 Memory estimate: 20.39 MiB, allocs estimate: 130.

kfolds without setdiff

julia> @benchmark kfolds(Int(1e5), 5) seconds = 600
BenchmarkTools.Trial: 10000 samples with 1 evaluation per sample.
 Range (min … max):  235.259 μs …   7.080 ms  ┊ GC (min … max):  0.00% … 88.45%
 Time  (median):     335.147 μs               ┊ GC (median):     0.00%
 Time  (mean ± σ):   454.221 μs ± 296.795 μs  ┊ GC (mean ± σ):  15.21% ± 17.02%

     ▆█▇▅▃▃▂▂▂▂▁                                ▁▁▁ ▁           ▁
  ▂▂███████████████▇█▇▆▆▆▆▇▆▆▄▅▆▇▅▆▇▇▇▇▇▇▇████████████▇▇▇▇▇▇▇▇▆ █
  235 μs        Histogram: log(frequency) by time       1.43 ms <

 Memory estimate: 3.05 MiB, allocs estimate: 25.

[CarloLucibello]

thanks!