cache: improve out-of-disk condition

When suspect SIGBUS happens, print helpful error and try to remove
the cache, so that the service might work again if auto-restarted.
Theoretically we could longjmp() out of the SIGBUS handler,
but that would be rather messy, so let the process die.