fix: validate fractions in get_dataset to prevent ZeroDivisionError

[CJstate]

What does this PR do?

Adds construction-time validation for fractions in get_dataset. Zero or negative values were accepted at construction time but caused a ZeroDivisionError (or silently produced negative selection sizes) during dataset loading, as reported in #6981.

The fix adds a check consistent with the existing validation for partial fractions and streaming:

if any(fraction < 0 for fraction in fractions):
    raise ValueError(f"All \`fraction\` values must be non-negative, got {fractions}")
if sum(fractions) <= 0:
    raise ValueError(f"Sum of \`fraction\` values must be positive, got {fractions} (sum={sum(fractions)})")

Also adds parametrized regression tests covering negative fractions and zero-sum fractions.

Fixes #6981

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? Please add a link to it if that's the case. (🐛 [Bug] get_dataset raises ZeroDivisionError for zero-sum fractions #6981)
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

AI writing disclosure

• AI-assisted: some parts were suggested or improved by AI, but the PR was written and reviewed by a human.

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag members/contributors who may be interested in your PR.

---

Note

Low Risk
Input validation and tests only; behavior for valid configs is unchanged and failures surface earlier with explicit errors.

Overview
get_dataset now validates dataset mixture fraction values before loading any data, so invalid configs fail fast with clear **ValueError**s instead of ZeroDivisionError or bad row selection during weighting.

When any dataset uses fraction, the mixture must still set it on all or none; additionally, every value must be non-negative and their sum must be positive. Regression tests cover negative fractions and all-zero sums (fixes #6981).

Separately, get_repetition_penalty_reward rejects ngram_size <= 0 at construction time (mirroring existing max_penalty checks), with parametrized tests for 0 and -1.

^{Reviewed by Cursor Bugbot for commit 8668b23. Bugbot is set up for automated code reviews on this repo. Configure here.}