Breaking the Tokenizer Barrier: On-Policy Distillation Across Model Families