RA3: Mid-Training with Temporal Action Abstractions for Faster Reinforcement Learning (RL) Post-Training in Code LLMs