fix(ragas): score empty retrieval contexts as zero

[Silu Panda (SiluPanda)]

Fixes #223.

ContextRelevancy now returns score 0 with an empty relevant-sentences list when the flattened context is empty. Python's sync and async paths and the TypeScript scorer short-circuit before calling the LLM judge. This prevents division by zero, NaN/null scores, or a hallucinated sentence turning empty retrieval into a perfect score.

The change leaves non-empty contexts and their existing score clamping unchanged. Tests cover "", [], and [""], assert that no judge request occurs, and check that the TypeScript result serializes as zero rather than null.

Validation:

• Before the fix: 6 new Python cases and 3 new TypeScript cases fail because they call the judge for empty context.
• Python: pytest py/autoevals/test_ragas.py -k 'not retrieval' -q gives 10 passed, 8 deselected.
• TypeScript: the ContextRelevancy and mocked AnswerCorrectness tests give 6 passed, 3 skipped.
• pnpm run build passes for CJS, ESM, and declarations.
• All configured pre-commit checks on the four changed files pass.

Live-provider tests were deliberately not run; validation used local/mocked tests with API credentials unset.

AI assistance: implemented and validated using Codex.