Don't wait for EOF in _handle_upload_pack_tail with side-band-64k

Don't wait for EOF in _handle_upload_pack_tail with side-band-64k

As of Sat, 12 Apr 2014 08:46:15 +0000, cloning, fetching or receiving a pack
from a GitHub repository with dulwich results in a hang. For example, the
command:

    bin/dulwich clone git@github.com:jelmer/dulwich.git

produces the following output:

    Reusing existing pack: 12675, done.
    Counting objects: 1408, done.
    Compressing objects: 100% (891/891), done.

and then hangs indefinitely.

This commit and the next one aim to fix that.

The indefinite hang is because _handle_upload_pack_tail waits for an EOF from
the server that never arrives. With canonical Git, both the client and the
server disconnect after the fetch is completed. With canonical Git servers,
Dulwich works fine since the server will disconnect. However, GitHub seems to be
running a non-canonical implementation of Git, which waits for the client to
disconnect. That works fine with canonical Git clients but not with Dulwich.

Is this a bug in the server or the client? This is arguable, but it's reasonable
to assume that no matter how the server acts, the client should disconnect
without waiting for an EOF. Why? Quoting Git's
Documentation/technical/pack-protocol.txt,

    After reference and capabilities discovery, the client can decide to
    terminate the connection by sending a flush-pkt, telling the server it can
    now gracefully terminate, and disconnect, when it does not need any pack
    data. This can happen with the ls-remote command, and also can happen when
    the client already is up-to-date.

This indicates that in general for stateful transports (i.e. not HTTP), the
client is responsible the lifecycle of the connection. The protocol
documentation doesn't explicitly specify what happens after the client fetches a
pack, but I think it's logical to extrapolate similar behaviour.

This patch causes the Dulwich client to disconnect right after the fetch is
complete, which menas that the above clone command completes correctly. Clones
from canonical Git servers continue to work as well.

It is pretty hard to write a test for this patch because the canonical Git
server doesn't behave this way and we don't have access to the GitHub server.

This was originally reported as a bug in hg-git, which uses Dulwich: see
https://bitbucket.org/durin42/hg-git/issue/104/hangs-after-total. Thanks to
John S Long and Marco Molteni for debugging this issue over there.