Since the concept of _complete transfers_ is being split out into
the _export_ concept, we lose knowledge of completeness in the GC.
Instead of asking exporters for transfer statuses we just treat all
transfer as incomplete when GCing.
Coverage becomes the shortest path problem on graph where nodes are
offsets and edges are nodes' segments. Implement a simple Dijkstra
algorithm to find one.
Otherwise there are situations when it's not entirely clear if a
transfer is really ready to be assembled. Since the `size` field
in a filemeta is not required (and rightly so), we need client to
tell us the final transfer size at the end of the process.
Also synthesize a testcase to show why it's needed.
Also worth noting that right now `fin` packets require final size,
even if a client already told us the size through filemeta. The
latter is regarded as serving informational purposes only (which
means that, for example, it might differ from the final size, or
some tranfer progress might show >100% somewhere because of that).
Andrew Mayorov