docs: add policy ratio failure-mode table

[yaoyao12345666]

What does this PR do?

Adds the first concrete failure-mode table for policy-ratio and rollout-drift debugging. It distinguishes expected GRPO/GSPO unit-valued surrogate ratios from stale-policy and masking problems, points to the relevant metrics/files/tests, and links the table from the troubleshooting index.

This is documentation-only and does not change runtime behavior.

Related issue

Fixes #55

Type of change

• Documentation update

How was it tested?

• python -m sphinx -E -a -W --keep-going -b html docs .docs-build-55
• ruff check docs --no-cache

The strict Sphinx build completed without warnings.

Checklist

• The PR title summarizes the contribution.
• Linked the related issue in the description.
• Existing tests pass (pytest tests/ -k cpu) - not rerun for this documentation-only change on Windows.
• New documentation passed strict Sphinx validation.
• Described the validation commands and environment limitation.
• No public API or CLI behavior changes.