Knapsack / MaskedKnapsack GPU sort

[samtalki]

_partial_sort_negative! is O(m·k) insertion sort, currently CPU-only.
oracles affected: Knapsack, MaskedKnapsack. blocks GPU support for
budget-constrained problems.

• options: bitonic top-k via KernelAbstractions.jl, thrust::partial_sort
  (CUDA), MPS sort (Metal)
• only worth it when k is large; small k stays on CPU
• sparse vertex protocol must still return nzind / nzval

cross-vendor parity: pick a path that works across backends.