[Feature]: Batched argmax reduction operator

Add a batched operator that returns, per row along a configurable axis, the index of the maximum element and its value. This is the "best-class" reduction used everywhere in inference post-processing.

Use cases

• Detection: best class + score per anchor ([B, N, C] → [B, N]), the input to top-k.
• Classification: top-1 label + confidence.
• Segmentation: per-pixel class map ([B, H, W, C] → [B, H, W]).

Requirements

• Input scores tensor with a configurable reduction axis (must handle the class axis in both [B, N, C] and [B, C, N] layouts).
• Output two tensors: argmax indices (integer) and corresponding values.
• Dtypes: fp32 inputs; integer index output.
• GPU and CPU paths, following the existing operator/kernel-wrapper pattern.

Acceptance

• Correct results vs a CPU golden model across shapes, axes, and both layouts.
• Ties resolved deterministically (e.g. lowest index wins) and documented.

No file or test is named. Start by locating existing operator and kernel-wrapper implementations, including both CPU and GPU paths, then identify the test structure used for shape and axis coverage. Done means matching a CPU golden model across the requested layouts and axes, returning indices and values, with deterministic tie handling documented.