Couldn't you signal the deadlocked process based on monitoring pidfd of the crashed one, then break the deadlock in the interrupt handler?
I'm not saying it's impossible, but you're basically implementing a distributed, embedded version of the lock daemon. And as all distributed things, fault-tolerance is hard. It's easier in a dedicated process or in the kernel because they're centralized (and if the kernel panics all processes stop; that can be emulated in a daemon).