Implement a first pass at actors in the API. (#242)

* Implement actor field for tasks

* Implement actor management in local scheduler.

* initial python frontend for actors

* import actors on worker

* IPython code completion and tests

* prepare creating actors through local schedulers

* add actor id to PyTask

* submit actor calls to local scheduler

* starting to integrate

* simple fix

* Fixes from rebasing.

* more work on python actors

* Improve local scheduler actor handlers.

* Pass actor ID to local scheduler when connecting a client.

* first working version of actors

* fixing actors

* fix creating two copies of the same actor

* fix actors

* remove sleep

* get rid of export synchronization

* update

* insert actor methods into the queue in the right order

* remove print statements

* make it compile again after rebase

* Minor updates.

* fix python actor ids

* Pass actor_id to start_worker.

* add test

* Minor changes.

* Update actor tests.

* Temporary plan for import counter.

* Temporarily fix import counters.

* Fix some tests.

* Fixes.

* Make actor creation non-blocking.

* Fix test?

* Fix actors on Python 2.

* fix rare case.

* Fix python 2 test.

* More tests.

* Small fixes.

* Linting.

* Revert tensorflow version to 0.12.0 temporarily.

* Small fix.

* Enhance inheritance test.

src/common/test/test_common.h

@@ -22,8 +22,8 @@ static inline task_spec *example_task_spec_with_args(int64_t num_args,
   task_id parent_task_id = globally_unique_id();
   function_id func_id = globally_unique_id();
   task_spec *task =
-      start_construct_task_spec(NIL_ID, parent_task_id, 0, func_id, num_args,
-                                num_returns, arg_value_size);
+      start_construct_task_spec(NIL_ID, parent_task_id, 0, NIL_ACTOR_ID, 0,
+                                func_id, num_args, num_returns, arg_value_size);
   for (int64_t i = 0; i < num_args; ++i) {
     object_id arg_id;
     if (arg_ids == NULL) {