improper setfile member now causes segfault

[mterenzi]

Implements the ticket's agreed approach: a setfile line that can't be stat()ed or
loaded now aborts the reload, and mtbl_fileset_init() returns NULL with errno
set instead of coming up with a silently empty fileset.

Before / after, using an mtbl as a setfile:
Dumped 0 entries. (exit 0)
Segmentation fault (exit 139)

Three things worth considering:

1. dnstable doesn't check the NULL, so it segfaults in mtbl_source_get_prefix
   rather than saying anything pointing the user to the real issue.

2. setfile_updated() commits the mtime before the reload runs
   (libmy/my_fileset.c:63), so a failed reload records a revision it never applied
   and the next call returns true. Probably fine though?

3. fs_load() assigns errno after the fact because mtbl_reader_init() doesn't
   set it on the not-an-mtbl paths. Works, but setting it at the source in
   reader.c would be more robust. That's a much bigger change than what the ticket needs though.