Refactor cuda::launch to use CUkernel on CUDA 13+

Starting with CUDA 13 driver does not require the current device to match the stream's device. We should be able to remove the `__ensure_current_context` in launch for CUDA 13+, but we currently use `CUfunction` internally, which is device specific. We should be able to refactor the code to use CUkernel for CUDA 13+ compilation and avoid setting the current context.

This won't make it to 3.2, filing an issue to look into it for 3.3