Add DetConS examples

[saud5150]

Add training examples for DetConS based on Efficient Visual Pretraining with Contrastive Detection.

The transform (DetConSTransform), loss (DetConSLoss), and a masked pooling utility (pool_masked in lightly/models/utils.py) are already implemented. The forward pass retains spatial backbone feature maps, pools them per mask region, and passes the resulting (B, M, D) embeddings to DetConSLoss. For the example, the built-in spatial grid heuristic (DetConSTransform(grid_size=(5, 5))) avoids any external segmentation dependency — the paper shows a 5×5 grid is optimal.

Tasks

• Add examples/pytorch/detcon.py
• Add examples/pytorch_lightning/detcon.py
• Add examples/pytorch_lightning_distributed/detcon.py

Follow examples/pytorch/densecl.py as the structural template. I'd like to take this on.

[liopeer]

Note: You probably will need some additional dependency like scikit-image or so in order to generate the pseudo-masks!