DetAny3D runtime adapter

The worker input recipe and decoding interface are adapted from:
https://github.com/OpenDriveLab/DetAny3D
Revision: 10e484be0837d80aa33cff6ed95a9be834a3f794
Apache License 2.0. Source: deploy.py, wrap_model.py and train_utils.py.
LibreYOLO changes: typed-array worker boundary, CPU compatibility, original-image
coordinate mapping, canonical cuboid axes, input validation and CLI integration.

CPU attention compatibility implements the documented xFormers 0.0.16 API:
https://github.com/facebookresearch/xformers
Revision: 6f3c20f223c0f051e4b2dc744cc7591591ee04c4
BSD 3-Clause. Copyright (c) Facebook, Inc. and its affiliates.
The xFormers Python package is separately installed. No xFormers kernels are
bundled. Deformable attention reuses LibreYOLO's existing Apache-2.0 implementation.

The model implementation and checkpoint are not bundled. The separately
installed upstream runtime includes UniDepth under CC BY-NC 4.0:
https://github.com/lpiccinelli-eth/UniDepth
Revision checked: 8d8cfe4c7ee15297099983607febf0d4f32eb3d6
No UniDepth implementation was used to write or copied into this adapter.
This does not grant commercial-use rights to the external runtime.
The official full checkpoint is mirrored byte-for-byte at
LibreYOLO/LibreDetAny3D revision eeb89d1b1d37ec5a361b1ec79b26ea0620d9f4c3.
OpenDriveLab declares no per-object checkpoint license; LibreYOLO's maintainer
approved that mirror by treating the project's Apache-2.0 statement as the
redistribution basis for the parameters DetAny3D itself trained. This is
LibreYOLO's disclosed interpretation and not a clarification from the DetAny3D
authors. Upstream documents the checkpoint's depth branch as initialized from
UniDepth v2, so LibreYOLO applies CC BY-NC 4.0 to the mirrored file as the
strictest known term. The weights are non-commercial and are not covered by
LibreYOLO's MIT license.
