Blow away the panic mutex in favor of using a single atomic_cmpset() on a
panic_cpu shared variable. I used a simple atomic operation here instead
of a spin lock as it seemed to be excessive overhead. Also, this can avoid
recursive panics if, for example, witness is broken.