support tracking updated aggregated states

[yl-lisen]

This close #446

Please write user-readable short description of the changes:

• Add updates tracking data (tracking changes/ updates .etc)
• Unified versioned serialize and deserialize of aggregator
• Remove some useless implementation in Aggregator

[chenziliang]

Summary,

1. Always highly recommended move the refactor part to a different PR especially the current one is critical enough, that will make the PR / review process way more efficient
2. I didn't see the importance / necessity of breaking prepareBlockAndFill into 3 pieces so far. If we really do need, a separate PR will be way clearer.
3. We may be able to clean up duplicate semantics like AggregateStateType vs TrackingUpdatesType
4. We may be able to clean up no_more_keys in this PR or in future PR
5. Avoid template as much as possible unless it has real benefits and we will be blowing our binary size / compile time quickly and also harder to maintain / debug. Simple / direct code is usually better.
6. Prefer writeVar for serde since they save disk / network bandwidth and fast as well. Avoid using writeText since they are slow and big.
7. Let's start to comment more as much as possible. Current code base is kinda of slowly out of control regarding what each function does, what they mean. A bit too many branches / conditions to understand. Without documentation / comments, it will be very hard for different people to pick up.

[yl-lisen]

Summary,

1. Always highly recommended move the refactor part to a different PR especially the current one is critical enough, that will make the PR / review process way more efficient
2. I didn't see the importance / necessity of breaking prepareBlockAndFill into 3 pieces so far. If we really do need, a separate PR will be way clearer.
3. We may be able to clean up duplicate semantics like AggregateStateType vs TrackingUpdatesType
4. We may be able to clean up no_more_keys in this PR or in future PR
5. Avoid template as much as possible unless it has real benefits and we will be blowing our binary size / compile time quickly and also harder to maintain / debug. Simple / direct code is usually better.
6. Prefer writeVar for serde since they save disk / network bandwidth and fast as well. Avoid using writeText since they are slow and big.
7. Let's start to comment more as much as possible. Current code base is kinda of slowly out of control regarding what each function does, what they mean. A bit too many branches / conditions to understand. Without documentation / comments, it will be very hard for different people to pick up.

Thanks, these are great suggestions.
I'm so sorry, i've separated part of the implementation, but it's still too big, I will separate the refactor parts in next PR