Description

Add native sensor_msgs/msg/Image support to torch_conversions library.

This change introduces APIs to:

• Allocate initialized CPU- or CUDA-backed image messages.
• Create writable HWC tensor views over image buffers.
• Create cloned or zero-copy HWC tensor views from received images.
• Copy tensors into existing image messages while preserving headers and image metadata.
• Allocate and populate image messages directly from HWC tensors.
• Preserve row padding through tensor strides.

The conversion path maintains CUDA buffer lifetime and stream-event ordering without introducing device or stream synchronization. It supports packed, byte-oriented encodings and validates encodings,
buffer sizes, row strides, tensor shapes, and data types.

Is this user-facing behavior change?

Yes. Users can now convert native ROS sensor_msgs/msg/Image messages directly to and from PyTorch at::Tensor objects.

Images are represented as HWC tensors. Zero-copy input views are available by passing clone=false.

Did you use Generative AI?

OpenAI Codex (GPT-5) was used to draft the changes in this pull request.

Additional Information

The current implementation supports packed, 8-bit-per-channel encodings, including mono8, rgb8, bgr8, rgba8, Bayer8, YUV422, and 8UC/8SC encodings. Planar and non-byte encodings are rejected
with descriptive errors.

For zero-copy CPU input views, the source message must remain alive while the tensor is in use. CUDA views retain the appropriate buffer handles and stream-event dependencies for the tensor lifetime.