The Price of Reasoning: Cost-Quality Tradeoffs in Reinforcement Learning for Neural Machine Translation