[FEA] how to align rmm stream usage with raft handle

[mfoerste4]

Several APIs in pylibraft implicitly call into rmm, e.g. to create a DeviceBuffer for a ndarray. The standard behavior for rmm is to use the default stream. This causes issues with any application relying on multiple streams.

The most intuitive way for a users would be if all rmm calls would inherit the handle/stream from the parent raft thread. This could be done manually for each individual call, but maybe there is a way to automatize this?

CC @tfeher