T5gemma2 (#41834)
* Fix small bug in T5Gemma 1 in __init__
* Add t5gemma2 model & configurations.
* Add auto support
* Add test case.
* Add doctree.
* Update positional embeddings to match latest update.
* Style fix & add use of final_logit_softcapping for attributes check.
* Update tests and embedding design.
* Add t5gemma2 to image-text-to-text category.
* Add T5Gemma2 doc.
* remove unused imports.
* minor update following comments.
* minor style fixes.
* fix config.
* Update T5Gemma2 following Anton's comments:
1. Override _prepare_cache_for_generation to take care of cross-attention cache.
2. Move vision preprocessing from main model to encoder.
3. Clean and fix bugs in modular model.
* Add T5Gemma2VisionConfig.
* Minor updates.
* fix style
* re-structure vision encoder and minor fixes.
* fix parameter tying.
* remove several unnecessary codes and fix small bugs.
* update and fix init.
* Update weight tying and other minor changes.
* Skip `tie_word_embeddings` in config attributes check in T5Gemma2.
* minor fix.
* fix the inherence of t5gemma2decoderlayer
* sync to head.
* update decorator usage
* disable FA and Flex due to merged module behavior
* style
---------
Co-authored-by: vasqu <antonprogamer@gmail.com>
