Reduce memory used for collection/sorting in SegmentReducer in the SegmentProcessorFramework

[npawar]

We added this SegmentProcessorFramework: https://github.com/apache/incubator-pinot/blob/master/pinot-core/src/main/java/org/apache/pinot/core/segment/processing/framework/SegmentProcessorFramework.java
Here’s some background about the framework: https://docs.google.com/document/d/1-e_9aHQB4HXS38ONtofdxNvMsGmAoYfSnc2LP88MbIc/edit#heading=h.ths3qhhyyy7w

This framework takes M segments and converts them into N segments. It has 3 stages - segmentmapper, segmentreducer, and creation of segments. In the reducer stage, we are doing aggregations and sorting: https://github.com/apache/incubator-pinot/blob/master/pinot-core/src/main/java/org/apache/pinot/core/segment/processing/framework/SegmentReducer.java#L114
This collection of GenericRows sorting happens in-memory, and can get very expensive.

We want to change this to off-heap implementation. So the tasks would be:

1. serialize the generic row (not the default object serde, but a custom one).
2. Use a PinotDataBuffer to store those off-heap
3. sort the rows in the buffer
   Refer to OffHeapSingleTreeBuilder’s sortAndAggregateSegmentRecords method. It pretty much does a similar thing.

As a decent first step, we can also start by simply changing the intermediate record format from GenericRow to something else like Record. GenericRow stores all values in a map, so the column names get repeated with every record, which unnecessarily inflates the memory needed.

[npawar]

Closing this, as these issues are resolved after the PR above and Jackie's refactorings.