Add zero-shot baseline mode: --adapter_path none skips adapter loading, --no_think suppresses Qwen3 thinking"

Add evaluate_v3.py — stratified sampling, layer-aware max tokens, incremental saves, resume support