Make the cuda kernel execution more robust + add log

We need to add a check on context launch and buffer creation. In particular, we need to ensure enough memory is available (or at least log it otherwise).