Better to review by each commit, some modification is scattered among lots of files.

Some major improvement:

1. Only retrieve srcId/type/rank/dstId in collectEdgeProps if needed
2. Only add FilterNode and AggregateNode if necessary.
3. Add FLAGS_enable_multi_versions, there is a string copy if we want to filter the edge of different versions.
4. Allocate RowReader from stack instead of heap, similar to fix VertexHolder::getDefaultProp performance issue. nebula#2249. The difference is that we use RowReaderWrapper to wrap reader of v1 and v2.

The Benchmark will retrieve 1 tag and 1000 edges, almost the same as vesoft-inc/nebula#2270, as follows:

1. Normal processor of GetNeighbors, one vertex, no filter.
2. Normal processor of GetNeighbors, one vertex, with filter. filter_ratio could modify the ratio of edges which could pass the filter.
3. Only use a SingleEdgeNode to get 1000 edges.
4. Directly use kvstore to get 1 tag and 1000 edges.

After these tuning, the benchmark of 2.0 has ~20% improvement than 1.0.