Dependency resolution bug

Notes for Gabriele

@ray.remote
def f():
  return 1
  
@ray.remote
def g(x):
  return x + 1

x_id = f.remote()  # ({'CPU': 1}, args=[])
g.remote(x_id)  # ({'CPU': 1}, args=[x]) --> ({'CPU': 1}, args=[])

Currently:
- f gets sent to worker A, g pending dependency resolution
- receive f's reply ({x: 1})
    - OnWorkerIdle(worker A) - no tasks in the queue
    - task_finisher_->CompletePendingTask
      - add {x: 1} into a in-process memory store
      - dependency resolver inlines x into task g
      - dependency resolver calls the callback for g, which adds g to the task queue
- request worker for g again

What we want:
- f gets sent to worker A, g pending dependency resolution
- receive f's reply ({x: 1})
    - task_finisher_->CompletePendingTask
      - add {x: 1} into a in-process memory store
      - dependency resolver inlines x into task g
      - dependency resolver calls the callback for g, which adds g to the task queue
    - OnWorkerIdle(worker A) - g is in the task queue, so we can reuse worker A

Code:
- Main class for task submission: https://github.com/ray-project/ray/blob/master/src/ray/core_worker/transport/direct_task_transport.cc
- Dependency resolution: https://github.com/ray-project/ray/blob/master/src/ray/core_worker/transport/dependency_resolver.h
- Processing task reply and return values: https://github.com/ray-project/ray/blob/master/src/ray/core_worker/task_manager.h
- Unit tests: https://github.com/ray-project/ray/blob/master/src/ray/core_worker/test/direct_task_transport_test.cc