feat: add JSON mode and reasoning_effort support

[kolmogorov-quyet]

Summary

Add two new features to improve LLM output quality and control:

1. JSON Mode (response_format="json_object")

• Add response_format parameter to summarize() method
• Enable JSON mode for LLM rankers in retrieve.py:
  • _llm_rank_categories
  • _llm_rank_items
  • _llm_rank_resources
• Enable JSON mode for _preprocess_conversation in memorize.py
• Support in openai_sdk.py, wrapper.py, and http_client.py

2. Reasoning Effort (reasoning_effort)

• Add reasoning_effort field to LLMConfig in settings.py
• Accepts values: "low", "medium", "high", or None
• Default value is "high" for better reasoning quality
• Pass through openai_sdk.py to chat.completions.create()
• Wire up in service.py when creating LLM clients

Why these changes?

1. JSON Mode: Ensures consistent JSON output from LLMs, reducing parsing errors like "No JSON object found"

2. Reasoning Effort: Enables control over reasoning depth for compatible models (OpenAI o-series, Cerebras, etc.)

[sairin1202]

Hi kolmogorov-quyet, thanks for the contribution! One small suggestion: it might be better to make these arguments optional, since not all models support them. As it stands, passing them unconditionally can cause errors for some models, for example:
openai.BadRequestError: Error code: 400 - {'error': {'message': 'Unrecognized request argument supplied: reasoning_effort', 'type': 'invalid_request_error', 'param': None, 'code': None}