Add rolling KV cache hook

[gueraf]

Summary

• Add a Self Forcing/Wan-focused (yet extensible) rolling self-attention KV cache hook for autoregressive video generation.
• Scope this first upstream step to rolling self-attention caching; cross-attention caching and pinned/sink-frame retention are not included yet.
• Add tests for cache append/overwrite/window behavior, context isolation, CacheMixin wiring, and Wan attention hook selection.

Motivation

This is a scoped follow-up to #12773 and a first step toward #12600. The previous draft explored similar functionality but also included Krea-specific experiments and broader integration work. This PR keeps the reusable self-attention cache mechanism isolated so it can be reviewed independently.

As for practical use, we (https://odyssey.ml/) would like to rely on the Hugging Face Diffusers ecosystem to ship Self-Forcing-like models without having to ship many custom modules, ideally none.

Bitwise equivalence against the original Self-Forcing implementation, plus example videos, are tracked at:
https://github.com/gueraf/self-forcing-diffusers/releases/tag/parity-artifacts-long-equivalence-7chunks-20260428