Fix continuous batching for multimodal models (#44436)
* Fix continuous batching for multimodal models
* Add continuous batching test for VLM
* Fall back to regular generate for non-text-only models
* Add warning when falling back from continuous batching
---------
Co-authored-by: jiwon-jeong <jiwon-jeong@lotte.net>
Co-authored-by: Rémi Ouazan <83456801+remi-or@users.noreply.github.com>
