Failure to fsync on XDR stream close

[graydon]

As noted in #2123 and #1395, we are seeing more instances of "missing buckets" than we expect from mere misconfiguration (which sometimes causes it, but not always).

On investigation with a node operator, we discovered a state that is best understood to be caused by our failure to fsync bucket writes:

• Bucket went missing at moment of node crash (cloud provider killed host)
• SQL database matched public network state (postgres WAL integrity maintained)
• Once restored, subsequent buckets diverge with difference appearing just to be missing entries or updates from lower levels (i.e. some merge restarted against a zero-sized bucket)

There are three necessary fixes for this. This bug tracks one of them: start properly fsync()'ing when we close a bucket (or any XDR file) so we don't lose them even if the OS crashes.

The other two fixes are better diagnostics (tracked with #2123) and verification of hashes while reading buckets (#2203).