Performance-Aligned LLMs for Generating Fast Code