[#821] Unregister the handler when a handshake aborts after registration

[vharseko]

Problem

DataServerHandler.startFromRemoteDS() and both ReplicationServerHandler start paths register the handler in the replication server domain before finalizeStart(), which can still fail before the reader and writer threads are started (interrupt in session.waitForStartup(); on the RS side the window also spans receiveTopoInfoFromRS(), which throws IOException/DirectoryException). abortStart() closed the session and rolled back the generation id but never unregistered the handler — and nothing else ever removes it: the normal cleanup is the reader or writer noticing the dead session and calling stopServer(), and neither thread was running.

The stale entry:

• inflates the connected DS count advertised in ReplServerStartDSMsg, skewing the DS-side weight-based load balancing forever;
• is published to peers and DSs in every TopologyMsg;
• keeps resetGenerationIdIfPossible() from ever resetting the generation id;
• permanently refuses reconnection of that same DS with ERR_DUPLICATE_SERVER_ID on every retry.

Fix

abortStart() now asks the domain to unregister a handler that got as far as registering (ReplicationServerDomain.unregisterFailedHandshake()), running the same cleanup stopServer() would have done: removal from connectedDSs/connectedRSs, monitoring publisher stop when it was the last server, handler shutdown, resetGenerationIdIfPossible() and topology notification.

Two deliberate differences from stopServer():

• the removal only fires when the map still holds that very handler instance, so aborts that happen before registration (e.g. the duplicate server id rejection) cannot evict a legitimately connected server;
• the cleanup runs under the already held domain lock instead of stopServer()'s interruptible lock acquisition, which would silently skip the cleanup when the handshake was aborted by an interrupt — the very trigger being cleaned up after.

Tests

New HandshakeAbortRegistrationTest pins the contract in three scenarios: an abort after registration unregisters the DS handler (including generation id reset and unblocked reconnection of the same server id), the same for an RS handler, and an abort before registration leaves the legitimately connected handler untouched. Without the fix the two regression scenarios fail with "the aborted handshake left a dead DataServerHandler/ReplicationServerHandler registered"; with the fix all three pass, as does the neighbouring HandshakeAbortGenerationIdTest, which drives real aborted handshakes through the modified abortStart() path.

Fixes #821.