Add judge and metrics modules (Phase 2)

[pratyush618]

Summary

• Add agenteval-judge module with OpenAI and Anthropic LLM-as-judge providers, HTTP client with exponential backoff retry, and JSON-first response parsing
• Add agenteval-metrics module with all 7 P0 evaluation metrics: AnswerRelevancy, Faithfulness, Correctness (G-Eval), Hallucination, Toxicity, ToolSelectionAccuracy, TaskCompletion
• Add shared infrastructure: LLMJudgeMetric abstract base class, PromptTemplate classpath loader, JudgeModels static factory

Details

agenteval-judge (12 source files, 6 test files)

• JudgeConfig builder with apiKey, model, baseUrl, timeout, maxRetries, temperature
• HttpJudgeClient with exponential backoff + jitter, Retry-After header support
• JudgeResponseParser: 3-tier score extraction (clean JSON → regex JSON block → bare score)
• OpenAiJudgeModel (POST /v1/chat/completions, json_object response format)
• AnthropicJudgeModel (POST /v1/messages, x-api-key + anthropic-version headers)
• JudgeModels.openai() / .anthropic() factory with env var API key resolution
• Unchecked exception hierarchy: JudgeException, JudgeRateLimitException, JudgeTimeoutException

agenteval-metrics (10 source files, 6 prompt templates, 9 test files)

• LLMJudgeMetric abstract base with final evaluate() template method lifecycle
• PromptTemplate with classpath resource loading, {{variable}} substitution, caching
• 5 response quality metrics: AnswerRelevancy, Faithfulness, Correctness (G-Eval), Hallucination, Toxicity
• 2 agent metrics: ToolSelectionAccuracy (deterministic F1/LCS), TaskCompletion (LLM-as-judge)
• 6 prompt templates as classpath .txt resources

Other changes

• Root POM: added modules, mockito 5.14.2, inter-module dependency management
• spotbugs-exclude.xml: added exclusion for abstract constructor throw pattern

Test plan

• mvn clean install passes — 180 tests (71 core + 39 judge + 70 metrics), 0 failures
• All unit tests use mocked JudgeModel — no API keys needed for CI
• Checkstyle, SpotBugs, and -Werror all pass
• Verify deterministic ToolSelectionAccuracyMetric scoring with edge cases
• Verify prompt templates render correctly with all variable combinations