Minimize intermediate state transition for bundle manager

[candicegjing]

Fixed #2485

Copied from the above ref:
There are essentially two issues here:

1. The bundle manager keeps assigning bundles to a worker that is going to be terminated soon.
   1. a bundle that is assigned to a terminating worker will stay in the starting state for 5 minutes until it gets picked up by the bundle manager.
   2. after the starting bundle gets picked up by the bundle manager, the worker goes offline, but there is still a record of this offline worker in the database.
2. The bundle manager keeps assigning bundles to a worker that is offline. (worker is disconnected, but its record still exists in the database)

Based on the above summary, there are differences and overlap between 1 and 2:

1. overlap: 1.ii and 2.
   1. goal: minimize the number of bundles that will be assigned to an offline worker.
   2. approach: detect the offline worker before dispatching a bundle.
2. difference: 1.a
   1. goal: minimize the number of bundles that will be assigned to a terminating worker.
   2. approach: detect the terminating worker before dispatching a bundle.

This PR aims to fix 1.a: minimize the number of bundles that will be assigned to a worker which will be terminating soon.

In the PR, I added a new column in the worker table is_terminating (please bear with my poor English, feel free to suggest changes in the naming). Whenever the termination signal is sent, the worker will update the database with is_terminating=True in the earliest time. Then when dispatching bundles from bundle manager, always check for this field before starting a new bundle on a worker. This way, we minimize the number of bundles that will be sent to a worker that is going to be terminated soon.

1. Pros:
   1. avoid bundles getting stuck in starting state for 5 minutes (messed up with bundles' execution priority)
   1. the intermediate state transition between starting and staged
2. Cons:
   1. one more database call --> performance overhead