[#728] Reject TCP self-connects in replication connect paths

[vharseko]

Fixes #728.

Problem

AssuredReplicationServerTest.testSafeDataLevelOne fails intermittently on
CI during setup (e.g. ubuntu-latest/25 on PR #727's run, unrelated to it):
createFakeReplicationDomain:321 expected [true] but found [false], with the
real cause right above in the server log:

msgID=6 Replication Server failed to start : could not bind to the listen port : 65531.
Error : Address already in use

Root cause

TCP self-connect (simultaneous open), the same bug class as
ZOOKEEPER-2101 / KAFKA-1836 / JDK-8067207:

1. Test listen ports come from TestCaseUtils.findFreePorts() = bind(0),
   i.e. they live inside the OS ephemeral port range (the OS assigned
   65531 in the first place).
2. When an RS is shut down between data-provider iterations, its former peers
   (ReplicationBrokers of fake domains, other RSs) keep reconnecting to the
   vacated port in a retry loop.
3. Each connect() draws a local ephemeral port; on Linux, when the kernel
   picks the destination port itself, the connect succeeds via TCP
   simultaneous open — the socket is connected to itself and now occupies
   the very port the next RS is about to bind.
4. ReplicationServer.initialize() fails to bind, only logs the error, and
   the RS runs without a listener — the fake domain then cannot connect.

TIME_WAIT was ruled out: the JDK enables SO_REUSEADDR on ServerSocket by
default on POSIX platforms (verified empirically), so rebinding over
TIME_WAIT remnants already works. macOS/BSD refuse self-connects (EINVAL),
consistent with the flake being ubuntu-only.

Fix

Detect self-connected sockets right after connect() in both outgoing
replication connect paths and fail the attempt with ConnectException:

• StaticUtils.isSelfConnection(Socket) — local endpoint == remote endpoint.
  A legitimate established connection can never satisfy this, so false
  positives are impossible.
• ReplicationBroker.performPhaseOneHandshake() — the existing error
  handling closes the socket (releasing the stolen port) and retries.
• ReplicationServer.connect() — the existing failure path closes the socket
  and blacklists the address for a few connect iterations.

Tests

• TestStaticUtils.testIsSelfConnectionFalseForNormalConnection — both ends
  of a normal loopback connection are not self-connections.
• TestStaticUtils.testIsSelfConnectionTrueForSelfConnectedSocket — forces a
  deterministic self-connect (binds the local end to the very port being
  connected) and asserts detection; skipped on BSD-based stacks (macOS
  refuses explicit self-connects with EINVAL), runs on the Linux CI legs.
• GenerationIdTest (4/4) and TestStaticUtils (1755, 1 skip on macOS) green
  locally with the guard in place — normal broker/RS connects unaffected.

Follow-ups tracked in #728: initialize() swallowing the bind failure, and
the isAddressInUse() dummy-connect probe being self-connect-prone itself.