lwanming/depth-anything-v2-small

The model (onnx/model_fp16.onnx) is sourced from onnx-community/depth-anything-v2-small. It is a Depth-Anything-V2 (with a DINOv2 backbone) ONNX model.

To resolve the issue where the original model constructs 5D tensors in its fused-QKV attention, we used convert_5d_to_4d.py to convert it into an equivalent model that maintains 4D attention tensors for better compatibility.