By now, this approach is only used by CUDA.

- A new configuration option has been added to control whether we want to
  exploit the concurrent kernel execution feature and use N execution
  streams, or use just one.

- Changing task execution's instrumentation:
  - Extrae events are emitted using CUDA events and callbacks.
  - A couple of functions have been added to Nanox's API to fit this change.
  - instrumentation_api family has been increased to 1001.
  - enableWDClosingEvents() function has no effect from now on for GPU Device.
  - This also implies a change in application's generated code that will
    call these API functions (Mercurium's DeviceProvider class basically).

- Task number event (WorkDescriptor ID) is now emitted while the user code
  of GPU tasks is being executed.

- WD's deferred events are flushed just before the GPU task is run.

Closing burst events specifying type (using 0 in those events representing a value or an unspecifed value); close #882

This reverts commit 072ac697, reversing
changes made to 9a87a06f.

Solving conflict between (e Event) and (e nanos_err_t) in instrumentation API; Error come from exception first implementation; see #18

Simplifying instrumentation mechanism, reducing event size and removing some instrumentation calls; see #679

Implementing enable/disable instrumentation pure virtual functions. Implementing these services in extrae plugin using Extrae_shutdown() and Extrae_restart() functions; see #563

As kernel calls are asynchronous, the event was actually raised inside the task
outline, before the task had finished, giving an erroneous timestamp. Currently,
this event is raised after the call to cudaThreadSynchronize().