Add support for Gemma3 VLM architecture (#1601)

* Add gemma3 text-only text-generation unit test

* Add gemma3 modeling tests

* add gemma3 pre-processing

* Add Gemma3ForConditionalGeneration class

* update registry

* formatting

* test that we don't load sessions unnecessarily

* support hybrid multimodal & decoder-only loading

* Improve cross-architecture loading