[FEATURE] Unified `embed` Interfaces for Vision Transformer Models for MIM Pretrain Research

This feature request is related to challenges in Masked Image Modeling (MIM) pre-training using vision transformer models in `timm`. Currently, embedding and feature extraction are tightly coupled within `forward_features`, making it difficult to inject mask operations after initial embedding and positional encoding but before the transformer stages, which is a common MIM requirement. Researchers need to access embedded tokens for masking before passing them through subsequent transformer layers.

## Describe the solution you'd like
I propose a refactoring of all vision transformer models(e.g. vit, swin_transformer, etc.) to uniformly expose two distinct interfaces:

- `embed(self, x)`: This method should take the input tensor x (e.g., image patches) and return the embedded vectors, including positional encodings if applicable. The output should be ready for the transformer encoder stages.
- `forward_stages(self, x)`: This method should take the output from embed(x) (i.e., the embedded and position-encoded tokens) and pass them through the transformer encoder layers.

This separation would make the existing `forward_features(x)` effectively equivalent to `forward_stages(embed(x))`. This allows researchers to easily perform mask operations on the embedded tokens returned by `embed(x)` before passing them to `forward_stages(x)`, enabling flexible MIM pre-training experiments.

## Describe alternatives you've considered
I have considered alternative solutions, such as adding a mask parameter directly to the `forward_features` and `forward `methods, similar to `vision_transformer`'s implementation.

https://github.com/huggingface/pytorch-image-models/blob/a7c5368ba0c8713dc1c9a98cc83bf46ddd02b0a0/timm/models/vision_transformer.py#L933-L935

While seemingly straightforward, this approach presents drawbacks:

- **Function Signature:** It still alters the method's interface, even if default parameter values mitigate direct breakage.
- **Mask Value Flexibility:** More importantly, it would limit flexibility in how masked-out positions are handled, restricting whether masked token values can be learned (e.g., a learnable mask token) or simply set to zero. Separating `embed` and `forward_stages` provides full control.

## Additional context
If this refactoring aligns with the library's design, I would gladly contribute a Pull Request to implement these changes across relevant vision transformer models.

Start with the referenced implementation in timm/models/vision_transformer.py around lines 933-935, then inspect the ViT and Swin Transformer model paths named in the issue. Determine how embedding, positional encoding, and transformer stages are currently coupled. Done means relevant vision transformer models consistently expose embed(x) and forward_stages(x), with forward_features(x) equivalent to their composition.

Written by the indexing model from the issue text.