[C++] SliceBuffer on CudaBuffer should return CudaBuffer

[asfimport]

Currently SliceBuffer on a CudaBuffer returns a plain Buffer instance, which is dangerous for unsuspecting consumers.

Reporter: Antoine Pitrou / @pitrou

Related issues:

• [C++] Create a device abstraction (is related to)

_{Note: This issue was originally created as ARROW-2446. Please see the migration documentation for further details.}

[asfimport]

Wes McKinney / @wesm:
Buffer slicing may be a hot path in IPC reads, so we need to take care with performance if making any changes to how these mechanics work

[asfimport]

Antoine Pitrou / @pitrou:
The original motivation (the dangerous aspect of returning a plain Buffer) has disappeared with the device API work. This would now be more of a convenience improvement, and I'm not sure it's worth the complication of introducing virtual method calls for buffer slices.

[asfimport]

Wes McKinney / @wesm:
Agreed