[codex] Add EvoClaw benchmark inference

[xingyaoww]

Summary

• add an EvoClaw inference entrypoint that follows the OpenHands benchmarks Evaluation flow
• discover EvoClaw repo directories from --data-root and launch DockerDevWorkspace from each EvoClaw base image
• upload task queue/SRS materials into the agent-server workspace and run the standard OpenHands SDK Agent/Conversation loop
• emit git patches and conversation trajectories through the existing benchmark output writer

Motivation

This keeps the OpenHands agent implementation on the OpenHands benchmarks side instead of injecting an EvoClaw-owned SDK runner into an existing container. EvoClaw can be exercised through the same workspace + agent server model used by the benchmark suite.

Notes

• This is an inference harness; it does not reimplement EvoClaw's DAG grader in this PR.
• The local environment does not currently include an EvoClaw-data checkout with metadata.json instances, so validation here is static/entrypoint-level.

Validation

• uv run --no-sync python -m py_compile benchmarks/evoclaw/run_infer.py benchmarks/evoclaw/config.py
• UV_CACHE_DIR=/mnt/data/evocloud/.uv-cache uv run --no-project --with ruff ruff check benchmarks/evoclaw/run_infer.py benchmarks/evoclaw/config.py
• git diff --check