[ViLT] Refactor output handling to align with standardized patterns

[aman-coder03]

What does this PR do?

This PR refactors ViLT's output handling to align with the standardized patterns used across the codebase.

Key changes:

• Removes manual hidden_states/attentions propagation and passes output_attentions, output_hidden_states, and return_dict cleanly through the encoder call in ViltModel.forward
• Adds **kwargs forwarding to all child model self.vilt(...) calls (ViltForMaskedLM, ViltForQuestionAnswering, ViltForImageAndTextRetrieval, ViltForTokenClassification) so output flags are correctly propagated from the top-level forward call down to the base model
• Fixes ViltForTokenClassification to handle the inputs_embeds path correctly when computing text_input_size
• Fixes ViltForImagesAndTextClassification to return None instead of empty lists for hidden_states and attentions when not requested, ensuring correct output length counting

Why is this needed?

This aligns ViLT with the new output handling patterns introduced in #43979.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: vilt