Keep a pre-initialized FunctionCallInfoData in AggStatePerAggData, and
re-use that at each row instead of doing InitFunctionCallInfoData each
time.  This saves only half a dozen assignments and maybe some stack
manipulation, and yet that seems to be good for a percent or two of the
overall query run time for simple aggregates such as count(*).  The cost
is that the FunctionCallInfoData (which is about a kilobyte, on 64-bit
machines) stays allocated for the duration of the query instead of being
short-lived stack data.  But we're already paying an equivalent space cost
for each regular FuncExpr or OpExpr node, so I don't feel bad about paying
it for aggregate functions.  The code seems a little cleaner this way too,
since the number of things passed to advance_transition_function decreases.