…er at large n)

merge_sortperm_lowmem! carries a comparator that dereferences v[ix] and v[iy]
from global memory on every binary-search step inside the merge pass, making
the effective traffic O(n log²n). merge_sortperm! instead copies the keys into
shared memory alongside the indices so all comparisons stay in L1/shared memory.

Benchmarks on RTX 5080 (CUDA 13.2, Julia 1.12):
  n=2^18:  0.541 ms → 0.286 ms  (1.9×)
  n=2^20:  2.185 ms → 0.490 ms  (4.5×)
  n=2^22: 10.668 ms → 2.847 ms  (3.7×)
  n=2^24: 53.453 ms → 11.900 ms (4.5×)

sortperm! is now within 1.3× of plain sort! across all tested sizes.

The public temp kwarg is preserved: it maps to temp_ix in merge_sortperm!
(same semantics — a pre-allocated index swap buffer).

Tests: extend sortperm testset with full permutation-validity checks, 6 new
element types (Int16/UInt16/Int64/UInt64/Float64/UInt8), edge sizes (n=1..2049),
data-distribution coverage, comparator options, temp-reuse, exact Base.sortperm
match, and a merge sort stability check.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>