KernelBench: Can LLMs Write GPU Kernels? – Benchmark and Toolkit, Torch –> CUDA