Sort fix

[chriso]

This fixes a bug where the optimized sorting routines weren't used on ARM64 (where purego=true). It fixes another issue where the non-asm sorting routines were swapping items when they were equal.

I've added benchmarks for indirect sorting, and added tests for pivot selection to make sure we aren't triggering worst case quicksort behavior.

[tysonmote]

Big improvement in cachette/lsm's sorting benchmarks, which I rigged up to sort 100K items that are randomly ordered and ordered ascending and descending:

name                            old time/op  new time/op  delta
BufferSort_sortedAscending-10   24.5ms ± 2%   3.3ms ± 1%   ~     (p=0.100 n=3+3)
BufferSort_sortedDescending-10  16.7ms ± 2%   3.4ms ± 2%   ~     (p=0.100 n=3+3)
BufferSort_random-10            8.49ms ± 1%  8.61ms ± 1%   ~     (p=0.200 n=3+3)

[kalamay]

Nice catch!