Add error handing to cu calls

[nightduck]

Running sample chain with 4k topics and more than 13 nodes (or possibly any application that allocates a large amount of data) will crash on (cuMemCreate in create_cuda_ringbuf_allocator)[https://github.com/nightduck/hazcat/blob/21f28231e0bb14e063a5a6a92e557e70e952c955/hazcat_allocators/src/cuda_ringbuf_allocator.c#L114]. I theorize this creates a memory that prevents subsequent applications from running.