provide optimized APIs accepting GPU device handles or ordinals

[drossetti]

The current APIs have an extra overhead due to retrieving the GPU ordinal.
That could be optimized away if the users directly passed the GPU handle/ordinal corresponding to the stream.