Validate current device and CUB stream matches

All CUB algorithms accept a user provided stream argument. We assume in many places that the current CUDA device matches to where the stream is going. For example, when we ask for the PTX architecture or the available SMEM on the current device. If the current device is not the one where kernels will be launched via the user-provided stream, errors may happen, like launching a kernel with more dynamic SMEM than is available on the device to where the stream goes to.

We should add a check, at least an assertion, to all CUB dispatching codes that the stream launches kernels to the same device as the current CUDA device.