Partitioner creates copy when used in @spawn()

[wlruys]

This is an old bug discovered when Milinda was working on the QR App. Just documenting it now as we've starting talking about improving data partitioning.

Consider the following:

    mapper = LDeviceSequenceBlocked(nblocks, placement=gpu_list)
    A_blocked = mapper.partition_tensor(A) 
    @spawn(placement=A_blocked[i])

This will always cause an unnecessary copy of A_blocked[i] from the GPU to host when launching the task.

Maybe we need to introduce "( )" operators to distinguish when a copy is made or if it can be inferred otherwise.