Deadlock in argon's profile0_core/streams/test4341.obu

If profile0_core/streams/test4341.obu from the Argon test suite is decoded with 50 threads or more, it deadlocks, while it runs successfully with 49 threads. 50 threads is the number of threads where the logic switches from 7 to 8 frame threads - thus it's also reproducible by directly requesting 8 frame threads.

To reproduce:

$ dav1d -i profile0_core/streams/test4341.obu -o n.null --threads 8 --framedelay 8