Change type of AVG aggregates to double

[kgyrtkirk]

The sql standard is not very restrictive regarding this:

If AVG is specified and DT is exact numeric, then the declared type of the result is an implemen-
tation-defined exact numeric type with precision not less than the precision of DT and scale not
less than the scale of DT.

so; using the same type is also ok (without patch);
however the avg of 0 and 1 is 0 right now because of the retention of the integer typ

Postgres,MySql and Oracle and Drill seem to increase precision ; mssql returns 0
http://sqlfiddle.com/#!9/6f7248/1

I think we should also increase precision as its already calculated more precisely

note: the Drill windowing related test cases we want to run also contain resultsets for which AVG is more precise

This PR has:

• been self-reviewed.

[imply-cheddar]

This looks fine to me. It does run the risk of some change in that some queries will start returning floating points where they hadn't before. AVG() as an aggregator exists more for compatibility with other systems than for actual usage (it's almost always better to aggregate the count and value separately and divide them). So, as long as the floating points are aligned with what you would normally expect from other systems it seems good to me.