Previously when s3fs had dirty local data and an application opened a
second fd it would remove the stat cache entry, query S3 for the size
which was still zero, then reinitialize FdEntry with this incorrect
size. Instead flush local data to S3 so this query works. It is
possible that a more involved patch could do this with a less
heavyweight approach but this requires changing open. This does not
completely fix git clone but allows it to proceed further. Also make
some cache methods const-correct. References #839.