[Feature][WIP] Implement KoLeoLoss for DINOv2/v3

[d3tk]

Description

This pull request implements the KoLeo regulrizer from Sablayrolles et al., 2019 as a loss regularization for DINOv2 as described in Oquab et al., 2024.

This resolves #319

I referenced the KoLeo Loss implementations from DINOv2 and DINOv3 GitHubs

Changes

• Added KoLeoLoss class in stable_pretraining/losses/dino.py:
• Implements the efficient version using pairwise dot products (torch.mm) and the diagonal filling trick.

Updated DINOv2Loss:

• Added koleo_loss_weight argument (default 0.0 to maintain backward compatibility).
• Updated forward to accept student_cls_features and compute KoLeo loss if enabled.

Added Tests:

• Created stable_pretraining/tests/unit/test_koleo_loss.py.
• Includes unit tests for initialization, forward pass (shapes, basic logic), and integration with DINOv2Loss.

Checklist

• I have read the Contributing document.
• The documentation is up-to-date with the changes I made (check build artifacts).
• All tests passed, and additional code has been covered with new tests.
• I have added the PR to the RELEASES.rst file.