Ownership of gpu tasks

[devreal]

Description

The GPU task structure is allocated by the DSL but free'd by the device management thread. That breaks the ownership and prevents certain optimizations, like eliding the allocation by integrating the GPU task structure into a higher-level task object.

Describe the solution you'd like

The release of the memory holding the GPU task should be done by the DSL, which allocated it. Remove the free of that structure from the device task and move it into the release callback provided by the DSL (i.e., DTD and PTG).

Describe alternatives you've considered

Adding another callback to free the structure but that adds no value.

Additional context

[bosilca]

I agree, this has bugged me for a while but I was not able to find a clean solution. The problem is that we cannot free the gpu_task in task_release because we don't have the gpu_task there and because the task itself does not have a pointer back to the gpu_task.

We could make the gpu_task derive from parsec_object_t (which would give us access to a constructor/destructor) but the assumption there is that the object is automatically freed (after the destructors are called) when the refcount reaches zero. Not a great solution either.

[devreal]

Ahh, I forgot that the task doesn't point back to the gpu task. Maybe a callback on the chore would do? It's in a chore callback where we create the gpu task so we might as well have a callback to destroy it. There is no use for it for host-side tasks but I don't see that as an issue, it's just ignored.

[bosilca]

Look at #688. I'm trying to test it, but the new MPI detection sucks on all our machines.