sort/argsort/topk: index order among tied values differs from upstream #33

Found by #13. sort returns the same values as upstream, but the order of indices among exactly tied values differs. Top-p sampling sorts probabilities, so at the boundary a tie decides whether a token is in the nucleus. This changes which tokens can be sampled, not just display order.

Done when

• Measure upstream's cpu tie order for sort (stable=False and stable=True, ascending and descending, along every dim, f32/bf16/int), and for topk and argsort.
• Our cpu kernels reproduce it exactly, as golden cases with tie-heavy inputs (all-equal, two-valued, ±0.0, NaN).
• Nullified (replace with an unstable sort and see red). Gate green.