fix: validate dataset mixture fractions before loading

[behroozazarkhalili]

What does this PR do?

get_dataset normalizes mixture fraction values by their sum, so an all-zero mixture raises ZeroDivisionError after every dataset has already been downloaded, and a negative fraction is silently turned into a negative weight. The validation now runs before any load_dataset call: fractions must all be set or all unset (moved up from the post-load site), non-negative, and sum to a positive value. Individual zero fractions with a positive total stay allowed. The streaming guard moves up with it. A parametrized regression test pins each rejection and asserts that datasets.load_dataset was never called.

Fixes #6981

Credit

The approach comes from #6982 by @cananoo, which was closed on 2026-09-09; this PR carries it as a single commit of mine, rebased onto current main.

Verification

• tests/test_cli_utils.py -k fraction: 6 passed on the rebased branch (3 new cases, 3 existing).
• ruff check and ruff format clean on both files.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a GitHub issue? 🐛 [Bug] get_dataset raises ZeroDivisionError for zero-sum fractions #6981
• Did you make sure to update the documentation with your changes? The fraction docstring now states the constraint.
• Did you write any new necessary tests?

---

Note

Low Risk
Behavior change is limited to invalid fraction configs; valid mixtures behave the same, with earlier, clearer errors and no accidental downloads.

Overview
get_dataset now validates mixture fraction settings before any Hub/local load, so bad configs fail fast with clear ValueErrors instead of downloading everything and then hitting ZeroDivisionError (all zeros) or odd negative weights.

The pre-load checks cover: fractions must be set on all datasets or none (moved up from after load), each value must be non-negative, the sum must be > 0, and fraction + streaming is still rejected. Weight normalization reuses the precomputed total_fraction. The DatasetConfig.fraction docstring documents the non-negative / positive-sum rule.

Tests add a parametrized case for zero-sum and negative fractions and assert datasets.load_dataset is never called when validation fails.

^{Reviewed by Cursor Bugbot for commit ebf4e9c. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.