Make DiffusionGemma trainable (#46568)
* Add bos_token_id to DiffusionGemmaGenerationConfig
* Expose encoder hidden state and per-example self-conditioning in DiffusionGemma
* Disable gradient checkpointing for DiffusionGemma
* Use Seq2Seq outputs instead of custom output classes
* Revert "Use Seq2Seq outputs instead of custom output classes"
This reverts commit b2cd4ecbd5f12b1fb570e102feead361f6cae0df.