Registers a fp32, last-dim, contiguous topk override at aten::topk /
topk.values for K in {64, 128, 256, 512, 1024}. Fused radix-select +
bitonic sort, one CTA per row, single launch.
Two specialisations selected at compile time by a deterministic
constexpr keyed off torch.use_deterministic_algorithms: a faster
atomic-gather + ord-only sort path...