CPU crunching is done when Event loop is idle, but only for 0.1s to allow HTTP requests about the processed data to not be delayed more than ~100mS.

Time slicing is a way for OSes to implement "parallel" processing on (single-core) CPUs, so it seems like you've reinvented that :-) So:

What is generally used in this situation. A second thread?

Yes, or second process, or some other way of running multiple processes in parallel; there are various ways to implement communication between them. Though I don't know much about AnyEvent, a quick search reveals e.g. AnyEvent::Fork, whose documentation also links to several other similar modules in the AnyEvent family.

Update: Turns out I showed you an example of this using Mojo::IOLoop::Subprocess almost exactly two years ago.