So.. make the requests serial. Just dump all the requests coming in over IPv6 into a queue, service that queue at a rate that's higher than the non-bot traffic requires but still low enough to be a problem for bots that can't self-limit to a reasonable rate. And of course, manage that queue intelligently so that you start dropping requests before you run out of RAM.
It’s hard to “throttle” a single isolated request from a lone IP.