Modalities not detected for some datasets using the Webdatasets format

I have found 2 examples of the modality detection code failing to recognize modalities in text and image datasets using the Webdataset format:

* https://huggingface.co/datasets/ProGamerGov/synthetic-dataset-1m-dalle3-high-quality-captions

* https://huggingface.co/datasets/CaptionEmporium/midjourney-niji-1m-llavanext

I'm not sure where in the modality detection code that things are failing: https://github.com/huggingface/dataset-viewer/blob/main/services/worker/src/worker/job_runners/dataset/modalities.py

Read services/worker/src/worker/job_runners/dataset/modalities.py first, then reproduce modality detection for the two linked Webdataset examples. Trace why their text and image modalities are missed and verify that both datasets are recognized correctly after the fix.

Written by the indexing model from the issue text.