Accept PyTorch tensors and JAX arrays as inputs out of the box

[Smirkey]

It would be great if powerboxes could accept common tensor / array types directly, especially:

• PyTorch tensors
• JAX arrays
• NumPy arrays

The goal is to let users call powerboxes APIs without having to manually convert inputs first.

For example, this should ideally work out of the box:

pb.iou(torch_boxes, other_torch_boxes)
pb.nms(torch_boxes, scores)
pb.area(jax_boxes)
pb.convert(boxes, in_fmt="xywh", out_fmt="xyxy")

A few important points i think:

• framework support should remain optional if possible
• format / dtype semantics should stay explicit and predictable
• behaviour should be consistent across supported input types

[gcanat]

Did you have something like this in mind ?

from functools import singledispatch
from importlib.util import find_spec

HAS_TORCH = find_spec("torch") is not None

# numpy typed function
@singledispatch
def rtree_rotated_nms(boxes: npt.NDArray, ...):
    # existing implem here

if HAS_TORCH:
    import torch
    @rtree_rotated_nms.register
    def _(boxes: torch.Tensor, scores: torch.Tensor, iou_threshold: float, score_threshold: float) -> torch.Tensor:
        device = boxes.device
        boxes_np = boxes.cpu().numpy()
        scores_np = scores.cpu().numpy().astype(np.float64)
        indices = rtree_rotated_nms(boxes_np, scores_np, iou_threshold=iou_threshold, score_threshold=score_threshold)
        return torch.from_numpy(indices).to(device=device)

[Smirkey]

mmh would be nice if rust side could use them natively as well
maybe with dlpack ?