Add LLM accuracy benchmark

[vtsaplin]

Description

Create a benchmark to measure and compare SQL generation accuracy across different LLM models (GPT-4o, Claude, Gemini, Ollama, etc.).

Requirements

• Define test cases: question → expected SQL/result pairs
• Add benchmark script that runs all cases against a specified model
• Report accuracy metrics (% correct queries, % correct results)

Example usage

python -m pytest tests/test_benchmark.py --model gpt-4o
python -m pytest tests/test_benchmark.py --model claude-3-5-sonnet-20241022

Output

Model: gpt-4o
Accuracy: 18/20 (90%)
Failed: "Show monthly trend" - wrong GROUP BY