Summary

• Fix LLMRerankerConfig to accept a nested llm dict config
• Fix LLMReranker to forward provider-specific settings (e.g. ollama_base_url) from nested config to LlmFactory

Fixes #3768

Problem

When configuring llm_reranker with a nested LLM provider like Ollama:

"reranker": {
    "provider": "llm_reranker",
    "config": {
        "llm": {
            "provider": "ollama",
            "config": {
                "model": "dengcao/Qwen3-Reranker-0.6B:F16",
                "ollama_base_url": "http://localhost:11434"
            }
        }
    }
}

The nested provider/model/URL were silently ignored because:

1. LLMRerankerConfig had no llm field — Pydantic silently dropped the dict
2. LLMReranker.__init__ hardcoded only 4 fields (model, temperature, max_tokens, api_key)
3. Provider-specific params like ollama_base_url were never forwarded
4. Everything fell back to provider="openai" / model="gpt-4o-mini", causing 401 errors

Changes

1. mem0/configs/rerankers/llm.py — Added llm: Optional[Dict[str, Any]] field to accept nested LLM config
2. mem0/reranker/llm_reranker.py — When config.llm is provided, use nested provider and config dict (including ollama_base_url, model, etc.) to create the LLM via LlmFactory.create(). Falls back to original flat-field behavior when no nested dict is present.

Test plan

• Configure llm_reranker with nested Ollama config — reranker uses Ollama instead of OpenAI
• Configure llm_reranker with flat config (no nested llm) — backward compatibility preserved
• Provider-specific settings like ollama_base_url forwarded correctly
• model, temperature, max_tokens, api_key defaults applied symmetrically in both paths