Towards Widening The Distillation Bottleneck for Reasoning Models