Replacement of "Sequence" feature with "List" breaks backward compatibility

PR #7634 replaced the Sequence feature with List in 4.0.0, so datasets saved with version 4.0.0 with that feature cannot be loaded by earlier versions. There is no clear option in 4.0.0 to use the legacy feature type to preserve backward compatibility.

Why is this a problem? I have a complex preprocessing and training pipeline dependent on 3.6.0; we manage a very large number of separate datasets that get concatenated during training. If just one of those datasets is saved with 4.0.0, they become unusable, and we have no way of "fixing" them. I can load them in 4.0.0 but I can't re-save with the legacy feature type, and I can't load it in 3.6.0 for obvious reasons.

Perhaps I'm missing something here, since the PR says that backward compatibility is preserved; if so, it's not obvious to me how.