Soft Specialists: $\alpha$-R\'enyi Ensembles for Uncertainty-Aware LLM Post-Training