feat(EvoAgentBench): add OpenAI-compatible embedding API for BrowseComp-Plus FAISS search

[CZH-THU]

Summary

BrowseComp-Plus previously loaded embedding models locally (HuggingFace + tevatron DenseModel), which requires significant GPU memory. This change adds optional OpenAI-compatible embedding API support so query vectors can be fetched from external services (vLLM, sglang, ModelScope, etc.) while keeping the FAISS index on CPU.

Changes:

• faiss_searcher.py: When embedding.api_base is configured, skip local model loading and call /v1/embeddings via the OpenAI client. Query text still uses the BrowseComp-Plus task_prefix for index compatibility. Includes ${ENV_VAR} resolution for secrets and _maybe_l2_normalize() to skip re-normalization when the API already returns unit vectors. Uses encoding_format="float" for ModelScope compatibility.
• start_mcp.py: Passes embedding.* fields from information_retrieval.yaml to MCP server CLI args.
• information_retrieval.yaml: Adds commented example config for the embedding API block.

Local model loading remains the default when embedding.api_base is not set.

Area

• Architecture method
• Benchmark
• Use case
• Documentation
• Developer experience
• CI, build, or release

Verification

# Code review: verified parameter flow start_mcp.py → mcp_server.py → FaissSearcher.parse_args()
# Verified OpenAI client call matches ModelScope embedding API example (base_url, model, encoding_format="float")
# Verified task_prefix is applied in API path (required for qwen3-embedding-8b index compatibility)
# Not run: full end-to-end benchmark (requires ModelScope token / local FAISS index / MCP server)

Checklist

• I kept the change scoped to the relevant area.
• I updated docs, examples, or setup notes when behavior changed.
• I added or updated tests when the change affects behavior.
• I did not commit secrets, .env files, dependency folders, or generated output.
• Active relative links in Markdown files resolve.

Notes for Reviewers

• Embedding logic lives in the MCP/FAISS layer (faiss_searcher.py), not in browsecomp_plus.py — the domain adapter only starts the MCP server via start_mcp.py.
• normalize must match how the FAISS index was built; default false aligns with the official qwen3-embedding-8b index. When normalize: true, already-normalized API responses are detected (L2 norm ≈ 1) and left unchanged.
• API model name must match the index (e.g. Qwen/Qwen3-Embedding-8B).
• Only searcher_type: faiss supports these new args; bm25 / custom searchers are unaffected.

---

Example config:

mcp_server:
  model_name: Qwen/Qwen3-Embedding-8B
  embedding:
    api_base: https://api-inference.modelscope.cn/v1
    api_key: ${EMBEDDING_API_KEY}
    model: Qwen/Qwen3-Embedding-8B
    normalize: false