Issue #1: tighten config handling and typed classification payload

[JasonZhangHub]

Summary

• simplify load_config() to stop swallowing unrecoverable config errors and use logging for successful loads
• centralize classification defaults (classification_temperature, classification_max_tokens, output_wrapper_tag) in load_config() so downstream code reads from one place
• replace the long return tuple annotation in load_classification_data() with a named tuple (ClassificationData)
• remove duplicated fallback branch in classify_items_with_model() now that defaults are applied at config load

Verification

• python3 -m py_compile openai_finetuner.py

Closes #1