switch from heap-based partialSort to quickselect

[niklasf]

Before (based on 13a0d52):

  HuffmanPgnBench.decode  avgt   40  4297.839 ± 63.417  us/op

After:

  HuffmanPgnBench.decode  avgt   40  3344.431 ± 15.417  us/op  # quickselect only <--

  HuffmanPgnBench.decode  avgt   20  3344.925 ± 31.764  us/op  # insertionSortCutoff = 3
  HuffmanPgnBench.decode  avgt   20  3345.941 ± 29.388  us/op  # insertionSortCutoff = 4
  HuffmanPgnBench.decode  avgt   40  3370.571 ± 24.539  us/op  # insertionSortCutoff = 5
  HuffmanPgnBench.decode  avgt   40  3365.226 ± 25.166  us/op  # insertionSortCutoff = 6
  HuffmanPgnBench.decode  avgt  120  3390.946 ± 17.816  us/op  # insertionSortCutoff = 7

  HuffmanPgnBench.decode  avgt   40  3415.760 ± 27.759  us/op  # sortCutoff = 4
  HuffmanPgnBench.decode  avgt   40  3431.564 ± 17.375  us/op  # sortCutoff = 5
  HuffmanPgnBench.decode  avgt   40  3437.750 ± 23.575  us/op  # sortCutoff = 6
  HuffmanPgnBench.decode  avgt   40  3430.231 ± 30.012  us/op  # sortCutoff = 7
  HuffmanPgnBench.decode  avgt   40  3430.460 ± 16.339  us/op  # sortCutoff = 8
  HuffmanPgnBench.decode  avgt   40  3504.320 ± 20.214  us/op  # sortCutoff = 10

[niklasf]

@isaacl Based on your hint that decoding can be optimized if we don't need the sorted list.

[niklasf]

Rebased and reevaluated. Switching from quickselect to sorting no longer seems to do much.

[isaacl]

fyi -- before the rebase/merge from master, I tested Arrays.sort and saw as a bit slower than your insertionSort for 7 elements.