Towards practical reinforcement learning for tokamak magnetic control