Before this PR, the sparse-indexer prefill step materialized a large fp32 logits matrix for every query row and visible key, causing massive memory usage and slowdowns when handling large key sets (up to 1M keys).

The dense prefill path computed and stored the full [rows, keys] logits matrix in fp32, requiring repeated kernel launches and large HBM traffic for writing and reading logits, which limited throughput and forced tradeoffs between memory budget and KV cache capacity.

This PR implements a single persistent CUDA kernel that fuses scoring and top-k selection for each row without materializing the full logits matrix, ranking compact 24-bit score codes instead and rescanning only rows that cannot be served by the gate exactly in fp32, drastically reducing memory traffic and kernel launches.

Fusing compute and selection steps into a single kernel and avoiding intermediate large buffers can significantly reduce memory bandwidth and improve throughput in large-scale GPU workloads.