Add a built-in llm-as-judge backend

[anilmurty]

Judge-scored benchmarks currently route through the DeepEval adapter. A lightweight built-in judge backend
(call a provider model directly with a scoring rubric) gives users a judge without the extra dependency. Add
it behind the existing Judge protocol (tjbench/judge.py) so --judge llm works.

Keep the verdict language honest: a judge reports a measured pass/fail against a rubric, never a guarantee.

Acceptance criteria: --judge llm --judge-model <spec> scores a judged benchmark via a provider model;
MockJudge path still works offline; tests added.