Support vision transformers

[guangy10]

Following PR #35124, we will add support for vision transformer models that are suitable for on-device deployment

[johnsutor]

What is the timeline for this? Is there any way I can contribute?

[guangy10]

@johnsutor I think we will prioritize to get the key features implemented first so that users can have a good experience to publish .pte models, and load cached one from hub. For connecting the exportable vision models to optimum, it will have to wait until after it. However, it would be super nice if you would like to contribute!

Here are what you need in order to contribute:

1. Register new tasks for the vision models under https://github.com/huggingface/optimum-executorch/tree/main/optimum/exporters/executorch/tasks, similar to causal_lm for "text-genertion" task.
2. Modify the existing xnnpack recipe as needed, e.g. https://github.com/huggingface/optimum-executorch/blob/main/optimum/exporters/executorch/recipes/xnnpack.py#L78, the lowering to .pte should just work
3. To run the .pte model using ExecuTorch runtime via pybind, you will need to implement a new modeling class similar to ExecuTorchModelForCausalLM for the vision tasks.

With step 1 & 2, you will be able to generate the pte models. Step 3, inference, can be done separately.