I am sending patch where kill was replaced by SendProcSignal. I tested it on pg_bench with 400 connections, and it works without problems.
I tested it under high load and it is works. But it is slower than just kill - under load the killing 400 connections needs about 1.5 sec.
Maybe for forced drop database we can increase max time limit to 10 seconds (maybe more (statement timeout)) ? It is granted so we wait just on sigterm handling. Other situations are finished by error. So in this case is very probable so waiting will be successful and then we can wait long time.