[mlir][NVGPU]: Fix op description of nvgpu.device_async_wait.

According to the NVIDIA documentation on `cp.async.wait_group`
(https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-async-wait-group-cp-async-wait-all),
the `numGroups` attribute in `nvgpu.device_async_wait` should give an upper
bound of pending async group count (instead of a lower bound) when the
executing thread can be unblocked.

Reviewed By: christopherbate

Differential Revision: https://reviews.llvm.org/D154046