Add 4 judge providers and TrajectoryOptimality metric

[pratyush618]

Summary

• Google Gemini judge provider — generateContent API with x-goog-api-key header auth and responseMimeType: application/json
• Azure OpenAI judge provider — Azure deployments endpoint with api-key header, configurable apiVersion
• Amazon Bedrock judge provider — AWS SigV4 request signing, Anthropic Messages API format for Claude models
• Custom HTTP judge provider — OpenAI-compatible endpoint for vLLM, LiteLLM, LocalAI, and other self-hosted servers
• TrajectoryOptimality metric — LLM-as-judge metric evaluating agent path efficiency, penalizing redundant tool calls, circular reasoning, and unnecessary steps; configurable maxSteps
• JudgeModels factory updated with google(), azure(), bedrock(), custom() methods
• 38 new tests, 632 total (all passing)

Test plan

• mvn clean install — all 17 modules build, 632 tests pass
• Checkstyle + SpotBugs: 0 violations
• Google: request URL, auth header, content/token extraction, empty candidates
• Azure: deployment URL with api-version, api-key header, default version fallback
• Bedrock: SigV4 auth headers, region extraction, session token, content parsing
• Custom: optional Bearer auth, OpenAI-compatible format, no-key mode
• TrajectoryOptimality: optimal/suboptimal scoring, maxSteps, validation, defaults