Scheduling a Parallel Sparse Direct Solver to Multiple GPUs