Improving Stochastic Action-Constrained Reinforcement Learning via Truncated Distributions