Add a with_workspaces function

[amontoison]

close #1767
@maleadt Can you review the PR?

[maleadt]

LGTM. Maybe we could unify both definitions by defining with_workspace in terms of with_workspaces and passing a 0 CPU size? Or a nothing size, returning nothing?

[amontoison]

LGTM. Maybe we could unify both definitions by defining with_workspace in terms of with_workspaces and passing a 0 CPU size? Or a nothing size, returning nothing?

I decided to pass a -1 CPU size. We can keep the same input type with this modification (Union{Integer,Function}) and we are sure to use two buffers even if a CUDA routine returns a size 0 for the CPU buffer.