ZOOKEEPER-4946: Bound the join in Login.shutdown()

[IvanKhanas]

Problem

Login.shutdown() called t.interrupt() once and then t.join() with no
timeout. The interrupt could be lost: Shell.runCommand() caught
InterruptedException from process.waitFor() and rethrew it as an
IOException without restoring the interrupt status. Login read that as a
failed kinit, retried, and went back to sleeping until the next scheduled
refresh, which is around 8 hours for a 10 hour ticket. The join waited all of it.

SendThread calls Login.shutdown() as its last action, and
ClientCnxn.close() joins SendThread without a timeout, so ZooKeeper.close()
could hang. NIOServerCnxnFactory and NettyServerCnxnFactory call
login.shutdown() during their own shutdown as well.

Fix

• Shell restores the interrupt status at both sites where it catches
  InterruptedException.
• Login records the request in a volatile flag instead of relying on a single
  interrupt, and checks it before running kinit and before entering
  reLogin().
• Login.shutdown() bounds the join with zookeeper.kerberos.shutdownTimeoutMs
  (5s default) and warns when the thread has to be abandoned. The thread is a
  daemon, so abandoning it does not keep the JVM alive.

Tests

• ShellTest is new: interrupts a thread inside Shell.execCommand and asserts
  the interrupt status survives. Runs in under 0.1s.
• KerberosTicketRenewalTest.shouldNotBlockForeverWhenRenewalThreadDoesNotExit
  parks the renewal thread in a call that ignores interrupts and asserts
  shutdown() still returns.

Both fail without the production change: the first on the interrupt status, the
second by timing out, which is the reported symptom. The three existing tests in
KerberosTicketRenewalTest are unaffected.