[codex] add semantic similarity rating

[lfiaschi]

Summary

Adds semantic-similarity rating (SSR) as an opt-in rating layer for AudienceKit rather than as a dataset adapter or model backend.

• Adds SSRAnchorSet, SSRResult, SemanticSimilarityRater, and SentenceTransformerEmbeddings.
• Adds SyntheticPanel.run_ssr_survey(...), which elicits free-text Likert answers, scores them into PMFs, and preserves both text and structured columns.
• Keeps embeddings injectable so hosted embeddings, local models, and tests can use the same scoring abstraction.
• Adds id_column, weight_column, and respondent metadata handling to SyntheticPanel for non-GSS panels.
• Documents SSR usage, custom anchors, optional ssr extra, and the standalone PyMC Labs SSR reference implementation.

Validation

• git diff origin/main..HEAD --check
• uv run --extra dev python -m pytest tests -q (28 passed)
• uv run python -m compileall audiencekit scripts
• uv build
• Read-only subagent review was run before PR creation; validation and custom-dataset metadata findings were fixed in 9948a61.

Notes

Direct Likert prompting remains available through run_survey(...). SSR is opt-in through run_ssr_survey(...) and should use anchors that match the Likert question being scored.