This way programming is more straightforward e.g. for case where we
first have to process identification and only then serve.

This is also more similar to how usual sockets work.

* master: (76 commits)
  Silently ignore answers to packets that aren't ignored on closed connection
  Do not process any packet for aborted connections
  Fix sorting of delayed events
  Do never parse any packet from aborted connection
  fixup! Add file descriptor and aborted flag to __repr__ of connections
  fixup! qa: add a basic assertion in Patch to detect when patched code changes
  qa: fix occasional deadlock when starting subprocesses in functional tests
  client: Add support for zodburi
  bug: on exist/crash, storage space for non-voted data may be leaked
  storage: fix commit activity when cells are discarded or when they become readable
  client: speed up cell sorting on read-access
  qa: new ConnectionFilter.retry()
  Fix race when tweak touches partitions that are being reported as replicated
  qa: add a basic assertion in Patch to detect when patched code changes
  Forbid read-accesses to cells that are actually non-readable
  Fix potential EMFILE when retrying to connect indefinitely
  The partition table must forget dropped nodes
  master: make sure that storage nodes have an up-to-date PT/NM when they're added
  In STOPPING cluster state, really wait for all transaction to be finished
  master: fix random crashes on shutdown when using several master nodes
  ...