* tapasco_jobs now define functions to attach memory transfers to jobs
  to be executed just-in-time before execution by the scheduler
* scheduler supports new constructs: if transfer length is marked > 0
  for an argument, it will allocate memory, transfer the data and
  updated the argument with the corresponding handle, as well as copy
  back and free the memory after return
* no top-level API is implemented yet, only tapasco_job level
* no malloc for local memories is implemented yet