Implement self-supervised contrastive pretraining

[piotrlaczkowski]

Describe your changes shortly

Implements a self-supervised contrastive pretraining stage inspired by ReConTab. This feature, configurable and activatable, uses an asymmetric autoencoder and contrastive loss to distill robust, invariant embeddings. It's fully integrated into all KDP feature processing pipelines and includes comprehensive tests and documentation.

Checklist

My code follows all rules defined in:

• pre-commits are passing !
• unite-tests are passing !
• code is well documented (docstrings in the corresponding format including examples of usage etc ...)
• code API is connected to the documentetion
• all requirements are added using poetry

---

_{Learn more about Cursor Agents}