Refactor model clients to use config-provided API key

[jenny-miromind]

This PR refactors how API keys are managed across the model clients. Previously, openai_client.py and anthropic_client.py directly read from environment variables (e.g., OPENAI_API_KEY, ANTHROPIC_API_KEY). This caused conflicts when using OpenAI’s official models as LLM-as-Judge while also switching to non-OpenAI but OpenAI-API-compatible model providers.

Key changes include:

• Removed environment variable lookups from clients: Clients now rely solely on the api_key passed via configuration.
• Added api_key to default configuration: A placeholder ("") ensures the field exists in all configs.
• Updated base client initialization: api_key is now read from configuration using cfg.llm.get("api_key").
• Adjusted benchmarking and tracing scripts: These scripts now explicitly pass API keys from environment variables into the configuration when needed, maintaining support for LLM-as-Judge workflows.

This ensures a clean separation of concerns:

• Clients only consume the keys explicitly provided by configuration.
• Environment variables are injected at the orchestration/script level, avoiding unintended coupling between different client usages.