[Feature] Optimize Prompts for Direct LLM Calls in Non-Agentic Workflows

[christso]

Description:

Problem: Users need to optimize system prompts for simple, direct LLM calls (e.g., via OpenAI API) for tasks like SQL code reviews or schema change validation, incorporating company-specific guidelines (e.g., attached PDF/YAML files for style rules).

Proposed Solution: Extend bbeval opt to support direct LLM optimization:

• Input: Base system prompt + attached guidelines file (e.g., --guidelines coding_rules.pdf).
• Use: BootstrapRS (≤5 trials) to tune prompt for accuracy on test tasks (e.g., "Review this SQL: [code]" → scored against expected feedback).
• Output: JSON prompt with enforced rules (e.g., required placeholders like {{guidelines}}, banned phrases).
• Integration: --mode direct-llm flag; validate via code_execution tool for SQL syntax.

Benefits: Enables quick tuning for non-agentic reviews; reduces manual prompt engineering by 80% on guideline adherence.