Add claim-support benchmark for transcript-grounded scoring

[dpaul0501]

This PR contains:

• New features
• Changes to dev-tools e.g. CI config / github tooling
• Docs
• Bug fixes
• Code refactor

What is the current behavior? (You can also link to an open issue here)

claim_support() is available as a transcript-aware model-graded scorer, but it only evaluates the final claim in a single pass.

What is the new behavior?

This PR adds an optional atomic-claim verification flow for claim_support() via decompose_claims=True.

• Splits compound claims into atomic subclaims
• Verifies each subclaim against the transcript/tool trace
• Aggregates subclaim verdicts into a final score
• Keeps the scorer model-agnostic, so any supported LLM backend can act as the judge

It also adds:

• a small benchmark-style regression test comparing baseline vs atomic verification
• a security-trace example task
• documentation updates so claim_support() appears in the built-in scorer list

Does this PR introduce a breaking change? (What changes might users need to make in their application due to this PR?)

No breaking change.

claim_support() keeps its existing behavior by default. Existing users do not need to change anything unless they want the new atomic-claim mode, in which case they can opt into decompose_claims=True.

Other information:

This is meant to make claim verification more robust for agent and security-trace evaluations, especially when answers contain multiple facts bundled into one response.

[dragonstyle]

Duplicate of #4166