Small Models Are (Still) Effective Cross-Domain Argument Extractors