Some linear probe experiments with dinov3_vith16plus to avoid overfitting on BACH

[NavneelSinghal]

This adds a few experiments for linear probing:

1. Pooling spatially adjacent patches to get bigger and fewer patches, which gives us effectively 5 + n/k^2 patches instead of 5 + n.
2. The above but without register tokens.
3. Tied low rank projections (random or learned) from d to d' for each patch token, then a combined linear probe, which gives us effectively 5 + n * d'/d patches.

With rudimentary experiments on BACH, the second method seems to perform better than a CLS-only linear probe, the first seems to roughly match it, and the third one underperforms it quite a bit.

[CLAassistant]

All committers have signed the CLA.