[enhance] export image encoder model to onnx

In some scenarios, the image encoder may need to export onnx. I attempted to export it and deploy it using onnxruntime, which showed a significant improvement in speed. Can I submit a PR to add this feature?

The issue names no files, tests, or entry points. First determine how the image encoder is exposed for inference and what export path is expected; done means a working ONNX export and ONNX Runtime deployment path, with the reported speed improvement verified.

Written by the indexing model from the issue text.