MAX function should not cast to double, but instead keep original type

When selecting a max of a large long, it truncates the result and loses precision.
For example:
`select MAX(1639054811930692679) from table` returns 1.63905481193069261E18, and loses precision.

Instead, the MAX function should return the original type. This should also apply to the other scalar aggregation functions.

Start by reproducing the large-long MAX query described in the issue and trace the MAX aggregation implementation in Apache Pinot. Check the other scalar aggregation functions as well; done means MAX and the related functions preserve the original long type without truncation or precision loss.

Written by the indexing model from the issue text.