feat: parallel multi-grader fan-out with consensus strategies

[christso]

Problem

Currently each target has a single grader_target for LLM-as-judge scoring. A single grader can have blind spots or biases that affect score reliability.

Proposal

Add grader_targets (plural) field for parallel grader fan-out:

- name: copilot-cli
  provider: copilot-cli
  grader_targets:
    - grader-openrouter
    - grader-gemini
  grader_strategy: consensus  # or: majority, any, all

Each grader scores independently in parallel, then the strategy aggregates:

Strategy	Behavior
consensus	All graders must agree (strictest)
majority	>50% of graders pass
any	At least one grader passes (most lenient)
all	Return all scores without aggregation (for analysis/comparison)

Result JSONL

When multiple graders are used, the result should include per-grader scores:

{
  "scores": [
    { "type": "llm-grader", "grader": "grader-openrouter", "score": 0.9 },
    { "type": "llm-grader", "grader": "grader-gemini", "score": 0.8 }
  ],
  "grader_strategy": "majority",
  "grader_agreement": 1.0
}

The grader_agreement field (0.0–1.0) measures inter-grader reliability.

Use cases

• Reduce grader bias: one LLM's blind spots covered by another
• Cross-provider validation: ensure scores aren't provider-dependent
• Confidence scoring: high agreement = high confidence in score
• A/B testing graders: compare grader quality before switching

Prior art

• Google ADK: multiple evaluator judges with voting
• LMSYS Chatbot Arena: multi-judge ranking
• No eval framework exposes this declaratively in YAML config yet

Backward compatibility

• grader_target (singular) continues to work unchanged
• grader_targets is optional; default strategy is majority
• When only one grader is specified in grader_targets, behaves identically to grader_target