transformers

c2fafabb - Fix VideoPrismForVideoClassification returning last_hidden_state as h… (#46830)

Fix VideoPrismForVideoClassification returning last_hidden_state as h… (#46830) * Fix VideoPrismForVideoClassification returning last_hidden_state as hidden_states ImageClassifierOutput.hidden_states must be tuple | None, but the model was passing vision_model_outputs.last_hidden_state (a raw tensor). Also adds the missing attentions field and updates tests to assert correct behavior. * Test recorded hidden_states/attentions for VideoPrismForVideoClassification Replace misleading skips with real tests: now that hidden_states/attentions are correctly forwarded from the vision backbone, add custom tests mirroring VideoPrismVisionModel that assert the spatial-then-temporal shapes when output_hidden_states/output_attentions are requested.