Vintix: Action Model via In-Context Reinforcement Learning