Add reusable LLM judge provider support for GPT-5.5 and DeepSeek V4 Pro

[taivu1998]

Summary

Adds a small reusable LLM judge provider layer for Evalchemy so new proof/rubric benchmarks can use GPT-5.5 and DeepSeek V4 Pro without copying provider-specific API code into each benchmark.

This PR adds:

• eval.judges.JudgeConfig with provider/model presets, redacted config metadata, env-var validation, CLI/YAML helpers, and stable config hashing.
• eval.judges.JudgeClient with OpenAI Responses API support for gpt-5.5 and OpenAI-compatible Chat Completions support for deepseek-v4-pro.
• Shared JudgeRequest, JudgeResult, and ProofJudgeResult schemas for structured rubric/proof grading.
• Evalchemy runner plumbing for judge CLI flags, nested YAML annotator: config, task constructor kwargs, output metadata, and resume fingerprints.
• Focused offline tests for config resolution, provider request payloads, schema validation, resume fingerprint behavior, YAML/argparse wiring, and task-manager constructor filtering.

Design notes

• Existing legacy annotator flows remain backward compatible. A flat legacy annotator_model such as gpt-4o-mini-2024-07-18 does not create a shared JudgeConfig or change resume metadata.
• The new judge layer is opt-in via GPT-5.5, DeepSeek model names, or explicit reusable judge provider settings.
• API keys are never stored in config metadata, output metadata, resume fingerprints, or tests; only env-var names are retained.
• OpenAI Responses calls pass store=False and use strict structured-output JSON schema.
• DeepSeek calls use response_format={"type":"json_object"} plus documented thinking-mode knobs through the OpenAI-compatible client.
• No existing benchmark internals are migrated in this PR; new proof/rubric benchmarks can opt in by accepting judge_config.

Validation

• python -m compileall -q eval/judges eval/eval.py eval/task.py eval/resume/wiring.py tests/judges
• python -m pytest tests/judges -q -rs -> 23 passed
• python -m pytest tests/resume/test_fingerprint_compute.py tests/resume/test_cli_wiring.py tests/test_passk.py -q -> 67 passed, 1 skipped
• python -m pytest tests -q -> 126 passed, 6 skipped
• git diff --check

No live provider calls were made; provider behavior is validated with mocked SDK clients and real request/response parsing code.