Intelligently blacklist scratch disks with physical I/O errors

I'm splitting this out from IMPALA-2079 since there are really two different parts to that.

This JIRA is to track process-wide blacklisting of scratch directories on disks that exhibit physical I/O errors that may indicate a failing disk.

We do not necessarily want to permanently blacklist disks, since that can result in daemons being in a bad state where they can't allocate any scratch space until they are restarted. Instead we should consider time-base blacklisting or some kind of administrative support to reenabled blacklisted disks.