fix(Qwen3VLProcessor): fix incorrect handling of return_metadata

[yiguanxianyu]

This PR fixes the following:

• ensures video metadata is correctly returned in Qwen3VLProcessor when return_metadata=True during processor.apply_chat_template().

[github-actions]

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

run-slow: qwen3_vl

[Rocketknight1]

cc @zucchini-nlp @molbap @yonigozlan

[zucchini-nlp]

Oh, interesting! I didn't know that the BatchFeature doesn't work well with non array-like structure. I don't think the current solution works well, it pushed us to update all models and the output type is different depending on the metadata

I recommend to fix it in def as_tensor() fn in BatchFeature, so that it doesn't try to convert everything to torch/np and instead checks the type of the input, whether it can be converted