feat(phoenix): support AgentV-authoritative LLM graders and rubrics

[christso]

Objective

Support llm-grader and rubrics in Phoenix experiments while preserving AgentV prompt/schema/scoring semantics.

Acceptance Signals

• Phoenix adapter runs AgentV's LLM/rubric grading path and logs the resulting score, verdict, assertions, evidence, and details into Phoenix evaluation metadata.
• Checklist rubric results preserve per-rubric assertions and evidence.
• Score-range rubric results preserve score, verdict, and details.
• LLM grader provider failures surface as clear evaluation errors.
• Unsupported custom prompt modes remain visible when they cannot safely run in adapter context.
• Support matrix and e2e notes are updated.

Implementation Notes

AgentV scoring should be authoritative first. Phoenix-native model evaluator templates can be considered later only after semantic differences are understood and documented.

Relevant files:

• packages/phoenix-adapter/src/evaluators/registry.ts
• packages/phoenix-adapter/src/evaluators/types.ts
• packages/phoenix-adapter/src/phoenix/run-experiment.ts
• packages/phoenix-adapter/test/evaluators/llm-grader.test.ts
• packages/phoenix-adapter/docs/support-matrix.md
• packages/phoenix-adapter/docs/e2e-verification.md

Non-goals

• Do not make Phoenix-native model evaluator scores the primary score for AgentV-authored rubrics in this issue.
• Do not add provider-specific config knobs for one-off grading behavior.