[ONNX export] Add depth-estimation w/ DPT+GLPN

[xenova]

What does this PR do?

This PR adds the ability to export DPT+GLPN models to ONNX for depth-estimation, unblocking huggingface/transformers.js#350.

... and it crosses DPT+GLPN off this list.

Running

optimum-cli export onnx -m hf-internal-testing/tiny-random-DPTModel o --task depth-estimation

now shows:

...
Validating ONNX model o/model.onnx...
        -[✓] ONNX model output names match reference model (predicted_depth)
        - Validating ONNX Model output "predicted_depth":
                -[✓] (2, 32, 32) matches (2, 32, 32)
                -[✓] all values close (atol: 1e-05)
The ONNX export succeeded and the exported model was saved at: o

Similarly,

optimum-cli export onnx -m hf-internal-testing/tiny-random-GLPNForDepthEstimation o

produces:

...
Validating ONNX model o/model.onnx...
        -[✓] ONNX model output names match reference model (predicted_depth)
        - Validating ONNX Model output "predicted_depth":
                -[✓] (2, 64, 64) matches (2, 64, 64)
                -[✓] all values close (atol: 1e-05)
The ONNX export succeeded and the exported model was saved at: o

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[xenova]

Note: The DPT conversion fails when a custom backbone (e.g., dinov2) is used. The associated PR was merged ~3 hours ago. This will need a closer look.

[fxmarty]

LGTM thank you! About the custom backbone, does this need to be fixed in this PR (aka, are large models using it)?

[xenova]

are large models using it

Nope, all the latest dpt-dinov2 models, highlighted in red below, have very few downloads (the most is 155, and the rest average around 10). The main two checkpoints from Intel, highlighted in green, have significantly more downloads.

So, I think it's okay to merge for now, then add support for these other models at a later date. I have no doubt that they will grow in popularity, considering they are released under the facebook org.

[fxmarty]

Sounds good!