Allow segment generation with a bounded memory

[xiangfu0]

Current segment creation requires unbounded memory as we use hashmap in AbstractColumnStatisticsCollector implementations.

This may fail minion tasks also future connectors like flink/spark/presto data sink.

The goal here is to allow configure a bounded memory size for segment creation.

@Jackie-Jiang

[Jackie-Jiang]

I think you are referring to the hash-set for deduplicating the values? We also stores the values in an array for sorting purpose.
We can use the off-heap data structure to reduce the memory usage, but that will also be inefficient.
I have plan to unify the segment creation for offline and real-time segment, and we might be able to borrow the off-heap data structure in real-time segments.