Adjust support for register tokens and add support for DINOv2 backbones with registers

[patricklabatut]

• Make applying positional encoding to register tokens optional (not done in DINOv2)
• Add missing initialization for register tokens
• Add support for DINOv2 backbones with registers (from Vision Transformers Need Registers)

Test: loaded DINOv2 backbones w/ registers with both timm and DINOv2's reference codebase and matched inference results to 1e-4 absolute tolerance (same as for DINOv2 backbones w/o registers) .

[HuggingFaceDocBuilderDev]

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

[patricklabatut]

looking into fixing the 10 broken tests...

[rwightman]

@patricklabatut so thanks for the quick turnaround on this, I already have a local draft of this and took a different approach, did not agree with their split of class vs reg embeddings and the ick logic it'd create so I just updated the weight adapter to make the reg ones no_embed_class compatible (model code does not change).

[patricklabatut]

@rwightman OK, shall I just drop this PR, then? If so, no problem, just looking forward to seeing the support merged.

[rwightman]

@patricklabatut yeah, I just have to verify the outputs are the same otherwise ready to push a PR, hopefully this weekend sometime.

[patricklabatut]

looking into fixing the 10 broken tests...

Pushed a potential fix for this only for reference.

[patricklabatut]

Closing this PR in favor of an alternative approach as discussed - thanks for working on this!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.