There is no need to react on every packet to mark a node as running. This
is updated on connection/disconnection, and it is sufficient.

Also, don't wait for client disconnection.
Also, factorise packet creation.

- Do not fetch data from outdated/discarded cells.
- Do not return more than transactions than requested by 'limit' parameter.
  Anyway, all results above this 'limit' could contain holes.

This fixes testStorageReconnectDuring{Transaction,Undo}Log unit tests.

The change in testStorageReconnectDuringTransactionLog fixes a bug in the test
itself. 'c.root()._p_serial' returned ZERO_TID.

Application.undo is also updated so that 'waitResponses' do not use
'setHandlerData' anymore.

This is another fix to commit 4720ba33.

Noticed because gitweb syntax colouring disliked it.

This fixes a regression in commit 4720ba33.

Inspired from "mdadm assemble --no-degraded" behaviour, which was brought
up during a discussion on linux-raid mailing list about split-brain
situation detection and avoidance.

PortAllocator is broken by design and the only proper solution to run a NEO
functional test without race condition with another software opening a random
port is to use ... unix sockets.

Keyword arguments are kept internally by the handler switcher, which passes
them to the handler when the peer answers.

This will be required by new replicator, so that 'setUnfinishedTIDList'
receives the list of outdated partitions at the time the
'AskUnfinishedTransactions' packet was sent.

This comment just described the whole purpose of the code, so there is no
XXX there.

If start() fails here, it fail in the next test too (and any test until
stop gets called).

For easier understanding of answerLastIDs' final else block.

Condition is true from the first iteration, as we just filled
unconnected_master_node_set.