KernelBench: Can LLMs Write Efficient GPU Kernels?