[FEA] Accept RAFT handle everywhere

[afender]

Context
There are many reasons why we may want a handle :

• communicator
• Keep track of library wide events (status)
• Counters (devices, streams, calls, … )
• Stream(s)
• Allocator
• Cuda library handles (expensive to create)

Todo:

• All multi-GPU should accept the handle from the beginning. It becomes pretty much required in the OPG paradigm. We need to pass the communicator so that we have a constant view of the multi-GPU / multi-process context.
• All single GPU C++ API should be upgraded to accept the handle.

[afender]

the handle is now in RAFT : rapidsai/raft#3

[afender]

The graph class has been updated and multi-GPU functions accept the handle.
We now need a consistent solution so that the handle content can be accessed from every feature.

Some of the challenges are:

• avoid undefined behavior when the graph has a different handle than the algo
• prevent recreating handle(s) as this can be costly

This task unblock stream support #957