Start GPU Dialect

    Define a new dialect related to GPU kernels.  Currently, it only contains a
    single operation for launching a kernel on a three-dimensional grid of thread
    blocks, following a model similar to that of CUDA.  In particular, the body of
    the kernel contains operations executed by each thread and uses region
    arguments to accept thread and block identifiers (similar to how the loop body
    region accepts the induction value).

--

PiperOrigin-RevId: 245713728