Brian/deep object detection

[brianzheng206]

• refactored deep_object_detection to work with any ONNX detection model (YOLO, DINO, Faster R-CNN, DETR, etc.)
• implemented a generic postprocessor automatically detects and adapts to model output format
• added configuratble pre-processing; multiple normalization schemes, resize methods (letterbox/resize/crop/pad), and RGB/BGR conversion

things that are supported:

• 3 channels, RGB/BGR only (no grayscale/multispectral)
• image data type is float32 only
• batch dimensions assumed to be dimension 0: [batch, channels=3, height, width] input and [batch, detections, features] output
• the auto-detect layout uses dim <= 8 heuristic (may fail on ambiguous shapes)
• batch size limit max 16

tested with yolov8m.onnx and camera_sync, publishing at ~30hz with tensorrt ep

[Edwardius]

BackendManager needs to not exist, these nodes should not need any backend-specific logic to function. Actually, the reason why is because this node is not a BASE CLASS of DeepNodeBase. Can you do that?

Also, please note that AI generated code is still YOUR code at the end of the day. Other people read it like its code written by you, so you should be more careful of when AI is creating code patterns that would never have happenned if you wrote it completely from scratch. Otherwise, your making others review AI code for you, and thats not fair.

[brianzheng206]

summary of changes:

• removed batching from the node entirely (don't know why i didnt clock that i was batching a batched message)
• deleted backend manager for good, inherits from deep node base to load backend plugins dynamically
• fixed the multi image msg names
• got rid of multi output for now
• got rid of auto-layout detection, just made a param for users to specify
• fixed the whole parameter situation...
• works with raw and compressed multi img msg

next things:

• deep bringup
• make multi output compatible

[Edwardius]

Tiny things now, otherwise LGTM!