Re-open files we get write errors writing to.

Do this rate-limitedly (currently, we require at least a minute between
reopens), in case something unexpected goes wrong.  It's most useful
for `files' like /dev/console (which I routinely see show EIO for
writes).