Make embedding and reranker batch sizes tunable via env vars

[justincasher]

Summary

• EmbeddingClient: add batch_size constructor param; falls back to LEAN_EXPLORE_EMBEDDING_BATCH_SIZE env var, then to the existing default of 8.
• RerankerClient: add batch_size constructor param that sets the default used by rerank(); falls back to LEAN_EXPLORE_RERANKER_BATCH_SIZE, then to the device-specific defaults (16 on CUDA, 32 on CPU).
• Lets deployments tune throughput vs. memory without code changes — raise on large-VRAM hardware for better throughput, lower on 8GB-VRAM boxes.
• Also fixes two pre-existing UP038 lint violations in scripts/generate_docs_data.py that were blocking any commit to main.