fix: raise a clear error when ValidSplit gets an IterableDataset

[aiedwardyi]

Closes #594.

ValidSplit needs len() and Subset, so an IterableDataset currently dies with an opaque TypeError about a missing length. It now raises a ValueError naming the cause and pointing at train_split=None.

Per the request on the issue that the error only be raised if really needed, the check sits in ValidSplit.__call__ rather than in NeuralNet - a custom train_split that tolerates an IterableDataset keeps working, and there is a test covering that.

Review fixes applied

• 961a68a: intercept the TypeError from get_len instead of checking isinstance up front, so an IterableDataset subclass that defines __len__ still splits normally; any other error is reraised untouched. Added a test for that case.
• 729cdcf: chain the original TypeError (from exc instead of from None) so the traceback keeps it; test asserts __cause__.

[BenjaminBossan]

Thanks for the PR. I made a suggestion, please check.

[BenjaminBossan]

Thanks for the update, just one question.

[BenjaminBossan]

Thanks for the update, LGTM.