Revisit Task and TaskCancel protocol

[sharpener6]

So in the current design, the task execution and cancel is not very ideal, which caused a few bugs:

Originally, in the very first version, all the tasks can be canceled on the worker side, even the one that is running, we achieved that by killing the process if that's the case, which means the ack for task cancel will be always success if task exists, and it doesn't need wait ack after task cancel happened, in this issue, we are going to formalize the task state changing, especially task cancel, so it can pave the road for following bugs:

1. balancing tasks use TaskCancel, without this proper ack mechanism, it cannot be solved properly Fixes balancing issues under heavy load. #51
2. when graph run get canceled, the graph tasks are not cleaned properly
3. the bug with NotFound issue Scheduler incorrectly returns a task not found error #45

I will upload the task state machine in following post under this issue

[yzard]