osd: fix lost objects
- track them properly
- consider rolling back to prior version vs strict EIO
- if a lost object had no prior_version, and we get a replayed request with a matching reqid, we should be thankful and reapply the write.
- probably we should do the same for an unfound or missing object, too!