MODNet portrait-matting model used by UseTool

Upstream project: https://github.com/ZHKKKe/MODNet
Upstream model: modnet.onnx, published by the upstream authors at
https://drive.google.com/file/d/1cgycTQlYXpTh26gB9FTnthE7AvruV8hd/view
Upstream file size: 25,888,640 bytes

The upstream README states that the code, models and demos in that repository
(excluding doc/gif) are released under Apache License 2.0. A copy of that
license is provided as /third-party/MODNet-LICENSE.txt.

UseTool modification:
- Converted constant floating-point tensors from float32 to float16.
- Kept model input and output types as float32.
- Conversion tool: onnxconverter-common 1.16.0, convert_float_to_float16,
  keep_io_types=true.
- Resulting file: /vendor/ai/modnet-portrait-matting-fp16.onnx
- Resulting size: 12,983,794 bytes
- SHA-256: 1990ab8035f2ccdb0af6c6866f6619342398b924eb58473e0bbcff2032bb7aab

No training, fine-tuning or dataset changes were performed. The model is used
only for portrait matting and is not presented as a general object-removal
model.