Part of the intent in commit fd1a421f was to allow SQL functions that are
declared to return VOID to contain anything, including an unrelated final
SELECT, the same as SQL-language procedures can.  However, the planner's
inlining logic didn't get that memo.  Fix it, and add some regression tests
covering this area, since evidently we had none.

In passing, clean up some typos in comments in create_function_3.sql,
and get rid of its none-too-safe assumption that DROP CASCADE notice
output is immutably ordered.

Per report from Prabhat Sahu.

Discussion: https://postgr.es/m/CANEvxPqxAj6nNHVcaXxpTeEFPmh24Whu+23emgjiuKrhJSct0A@mail.gmail.com