task dag enhancement.

[kyungjoo-kim]

There are missing features in Kokkos task dag capability regarding to CUDA space

• user interface to set cuda stack size;
• per team and per thread;
• user interface to set vector size currently 1x32 is hardwired.

I sent a patch to @hcedwar and this issue is for a tracking purpose.

[hcedwar]

@kyungjoo-kim : Integration testing fixes to CUDA back-end

The code is finally working. I need to have kokkos to be updated with the attached patch.
After we fixed capturing the member variables in a class, the error related with "std::move" does not appear anymore and I removed the modification.
The patch includes.

• memory pool kokkos inline decoration
• task team member need to have team_barrier() and team_broadcast.
• I reduced the stack size to default 1024. I tested it and it does not need larger stack anymore (but who knows someday I need it again; so I do not completely remove the interface in the kokkos.

kj-task.patch.txt

[kyungjoo-kim]

tacho.patch.zip

Oh.... the previous patch miss one thing related to team broadcast. Sorry that I am late for putting this patch.