Simple mechanism to run Triton kernels on PyTorch for debugging purpose (upstream from Kernl). Todo: - random grid iteration - support of atomic ops - more unit tests - cover new APIs?
