The explained segfault problem could be solved by cloning perl in the c callback handler as follows
Not really. The problem is that the interpreter you're trying to clone is owned by another thread. That thread might be actively executing, causing the state of that interpreter structure (and all the data hanging off it) to be changing at the same time that a second thread is trying to clone it.
It all depends on what the original perl thread is doing. If it just calls a 'start event loop' function in mpv which suspends the thread and lets mpv's own threads take over to do all the active work and callbacks, then you could just set my_perl to point to the original perl's interpeter and let the mpv thread update that. If multiple mpv threads can call callbacks simultaneously, then you'd need either to have a lock so that only one thread can use the interpreter at a time, or you'd need to create enough perl threads such that there is a big enough pool of interpreters to be borrowed by all active mpv threads. Of course with multiple interpreters you'd have the issue of only one interprter getting updated, not necessarily the one you want.
Or you could have some sort of queue system whereby mpv callbacks just put requests into queue which are popped and processed by one or more perl threads.
But without having some details of what the mpv library does and how, this is mainly speculation.