HashAggregate

[GindaChen]

This issue extends #57 to make sure HashAggregate delivers the following:

• Replace the std::unordered_map hash tables to phmap (ideally) for better performance.
• Add a toggle in AggregateStrategy to flip between HashAggregate and Aggregate. We can also introduce this toggle in ssb_workload.cc and ssb_workload_lip.cc to compare the performance.
• Handle hash conflict more elegantly. (Using phmap and custom hash_combine function can eliminate the problem).
• Design an abstraction over the hash table, and remove some redundant type-checking codes of the arrow arrays.
• Refactor arrow_compute_wrapper and add some helpers to optimize the array construction.
• Eliminate most data copy between std:: and arrow::.
• No memory leak. Free all memory after the Finish().
• Make code readable: use less auto

[GindaChen]

See #67 for primer benchmark result.