Gracefully handle OOM in Barrier (NVIDIA#483)

CUDSS does not throw on OOM and even though we check for the return status. There was a sticky cuda error on any successive CUDA API call. A workaround is to throw if `cudaMallocAsync` fails in the device mem handler and catch that error in the solve method.

Authors:
  - Hugo Linsenmaier (https://github.com/hlinsen)

Approvers:
  - Chris Maes (https://github.com/chris-maes)

URL: NVIDIA#483

Author: hlinsen