[Feature]: Batched top-k selection operator

Add a batched operator that returns the k largest elements' indices and values along a configurable axis, optionally sorted and gated by a score threshold. This replaces IoU-NMS in end-to-end detectors, which select a bounded top-k of candidates directly.

Use cases

• NMS-free detection: select top-K candidates by confidence (YOLO26/YOLOv10 one-to-one head; DETR top-k over queries).
• Classification: top-5 predictions.

Requirements

• Input scores tensor + configurable axis, K, and optional score threshold (entries below threshold excluded / padded).
• Output top-k indices and values per batch ([B, N] → [B, K]); sorted-descending option.
• Define behavior when valid entries < K (padding / count output).
• Dtypes: fp32 inputs; integer index output.
• GPU and CPU paths, following the existing operator/kernel-wrapper pattern.

Acceptance

• Correct top-k (indices + values) vs a CPU golden model across shapes, K values, and the threshold path.
• Deterministic tie/ordering behavior, documented.

Start by surveying the existing operator/kernel-wrapper pattern and its CPU and GPU implementations. Validate the operator against a CPU golden model across shapes, K values, threshold cases, and ties; done means indices and values, padding or counts for fewer valid entries, deterministic ordering, and behavior are documented.