Implement getting the cluster nodes list from nodes

[msiomkin]

Periodically call a user-defined Lua function on a node to obtain or refresh the full cluster nodes list.

Resolves: #134

[Totktonada]

BTW, there is the task about distros updating (#123). I have pinged Alexander about that.

[Totktonada]

I decided to skip review ping-pong and reworked some parts of the patch where I guess I can improve things. I have made the following changes:

• Fixed all flake8 warnings in mesh_connection.py and test_mesh.py: trailing whitespaces and long lines.
• Prevent a user provided addrs argument from changing: copy it before inserting host+port.
• Added warnings for parse_uri() and other cases like a network error during a call to a discovery function.
• Removed _opt_reconnect() before a call to a discovery function, but give a warning if something went wrong during the call.
• Restrict validation (say, catch ':1234' or 'localhost:65536'), extract it from RoundRobinStrategy; reworked to return warning messages as values to use as warnings or error (see the following bullet).
• Raise a configuration error if user provided addresses does not pass validation. Give warnings as before for cluster discovery function result.
• Added test cases for wrong cluster discovery function results.
• Give a configuration error for empty address list even with connect_now=False. It is needed to keep self.{host,port} consistent with self.strategy.pos and so simplify the code: say, we don't need to explicitly get a next address in connect(), only in _opt_reconnect().
• Reverted cycle usage in RoundRobinStrategy to implement self.pos adjusting properly (when a current address is in the new address list).
• Reverted self.connected = False in connection.py::close(). I looked at the git history and it is intended to be the internal field. Added test case for connect() after close() in PR #144. If we want to provide the way to determine whether a connection is alive we need to expose a function. But maybe ping() should be used for this.
• Fixed the bug when self.strategy.getnext() gives currently used address: this is why +1 was needed in the loop around super(...)._opt_reconnect() + getnext() in MeshConnection._opt_reconnect().
• Re-raise a last error in MeshConnection._opt_reconnect() (before it raises NetworkError class).
• Reworked test_mesh.py a bit: added / enhanced comments, cropped extra arguments where it is appropriate, some taste changes.
• Expanded the commit message.

Maybe I missed to mention something in the list.

@robinhood23rus Can you please give me feedback about the new version of the patch (force-pushed here).

[Totktonada]

Reorganized parse_uri() to get rid of asserts and make it obvious that its conditions are always met.

[denis-ignatenko]

Changes ok. Approve