Fix high memory usage during gemma4 image prefill (#362)

* Fix Gemma 4 image prefill chunking

* Simplify Gemma 4 image prefill invariants

* Align Gemma 4 prefill to cache boundaries

* Use image spans for Gemma 4 mask dispatch

* Clarify Gemma 4 image-run chunking name

* Clarify Gemma 4 image token terminology

* Explain Gemma 4 text-slice token types

* Remove unreachable Gemma 4 video check

* Document Gemma 4 image prefill sections

* Hoist Gemma 4 cache mask import

* Document Gemma 4 mask patch source

* Use token types for Gemma 4 image runs

* Keep touching Gemma 4 image sections separate

* Schedule Gemma 4 prefill by image runs

* Validate Gemma 4 aligned prefill boundaries

* Guard absent Gemma 4 sliding masks