fix: preserve multiple candidates in multimodal pipelines (#50198)
* fix: preserve multiple candidates in multimodal pipelines
* fix: address multimodal pipeline review feedback
* test(pipelines): verify real Omni audio generation
---------
Co-authored-by: BlazeSR10hawk <175792037+BlazeSR10hawk@users.noreply.github.com>
Co-authored-by: BlazeSR10hawk <BlazeSR10hawk@users.noreply.github.com>