New SSL method: Franca (Nested Matryoshka Clustering)

Nobody has claimed this yet.

## Overview

Franca (Venkataramanan et al., valeoai) looks like a good fit for lightly's DINO/DINOv2 family.

- **Paper:** https://arxiv.org/abs/2507.14137
- **Code:** https://github.com/valeoai/Franca (Apache-2.0)

## Franca contributions

It's a fully open DINOv2-based model (code, data, weights). It keeps the DINOv2 setup and adds:

- **Nested "Matryoshka" clustering**: the clustering head runs at several nested embedding sizes at once, giving coarse-to-fine assignments.
- **Cyclic-shift masking** for the iBOT branch.
- A post-hoc **positional-disentanglement head (RASA)**.

On ImageNet-21K, and separately a LAION subset, it reports matching or surpassing DINOv2, CLIP, and SigLIPv2.

## Fit with lightly

Most of it is the DINOv2 recipe already in lightly, so the main new pieces would be the Matryoshka head and loss, plus the cyclic-shift masking. I think we should leave out RASA as it's post-hoc (?)

Would it be of interest to include the method in lightly? If so I can try to implement it after agreeing on the implementation plan.

Start by reviewing lightly's existing DINO/DINOv2 implementation, then compare it with the Franca paper and the linked valeoai/Franca code. Clarify the implementation plan for the nested Matryoshka clustering head and loss and cyclic-shift masking, including whether RASA is excluded; done means the scope and design are agreed before coding.

Written by the indexing model from the issue text.