Randomize candidate dataset order

fix: reduce templates to country, region & county for mvp

Lowercase subtype names, fix question type hints

Fix: No pairs are created for mixed queries

enh: Add templates to handle queries like subregion, region

prompt: Add notes on how to pick places from query better

fix: mixed templates are not generated because of stale parquet files

chore: clean sql generation, use conversation format, move prompts from user to system

Add finetune instructions for gemma

FEAT: Add to create SLM training data