GPU Kernels, Part 2: Teaching a 27B Model to Write Triton Kernels with SFT and a GPU Gym