Raise instead of hanging on a probability-0 source without replacement

[Arthur031221]

Anyone who passes probabilities with a 0 weight together with stopping_strategy="all_exhausted_without_replacement" to interleave_datasets on map-style datasets gets a call that never returns: the source with probability 0 is never drawn, so it is never marked exhausted and the sampling loop keeps drawing forever.

from datasets import Dataset, interleave_datasets

a = Dataset.from_dict({"x": [0, 1, 2]})
b = Dataset.from_dict({"x": [10, 11]})
interleave_datasets([a, b], probabilities=[1.0, 0.0], seed=42, stopping_strategy="all_exhausted_without_replacement")

On main (fa995bd) and on 5.0.1 this does not return (killed by timeout 60). #8318 added a clear ValueError for the same case under all_exhausted, but that check sits in the branch that handles only first_exhausted and all_exhausted, so the without replacement loop still spins.

This adds the same check at the top of the without replacement branch. The call above now raises:

ValueError: interleave_datasets with stopping_strategy="all_exhausted_without_replacement" cannot exhaust a dataset that is never sampled; datasets at indices [1] have probability 0.

The first_exhausted and all_exhausted code paths are not touched.

I went with raising rather than treating probability-0 sources as already exhausted, to match what #8318 chose for all_exhausted (its test expects a ValueError even when the probability-0 source is empty). If you would rather skip such sources, initialising is_exhausted from probabilities == 0 is a one-line alternative (the call above then returns [0, 1, 2]) and I can switch to it. #8627 fixes empty sources in this same branch; that is a separate condition from this one.

Test: test_interleave_datasets_probabilities_zero_probability_all_exhausted_raises is now parametrized over both strategies and matches the message. With the source change reverted, the two new all_exhausted_without_replacement cases time out (pytest-timeout at 60s) and the two all_exhausted cases pass; with the change all four pass.

The streaming path (IterableDataset) has the same stall under all_exhausted_without_replacement: it yields the three rows of the first source and then blocks. I kept this PR to the map-style path.