Sending cluster time to an arbiter with auth enabled results in an error

Commands including a $clusterTime sent to an arbiter result in an error when auth is enabled on the server.

I reproduced locally by creating a three node (PSA) replica set with auth enabled using mlaunch:

mlaunch init --replicaset --arbiter --nodes=2 --auth --username user --password password --name replicaSet --setParameter enableTestCommands=1 --verbose --dir repro

The error can be reproduced by sending a command to the arbiter with a $clusterTime obtained from another node. Example with the shell:

var cmd = {

  "isMaster": 1,

  "$clusterTime" : {

    "clusterTime" : Timestamp(1591812742, 1),

    "signature" : {

        "hash" : BinData(0,"AAAAAAAAAAAAAAAAAAAAAAAAAAA="),

        "keyId" : NumberLong(0)

    }

  }

}

db.runCommand(cmd)

Replies with:

{

	"ok" : 0,

	"errmsg" : "aborting keys cache refresh because node is shutting down",

	"code" : 91,

	"codeName" : "ShutdownInProgress"

}

Interestingly, this occurs without authenticating the connection, but I cannot reproduce this if the server is started without auth enabled.

Behavior appears in server version 4.5.0-1606-gbd74e90, but not 4.5.0-1602-g5c6e1e5.

The most relevant looking server commit is 7074dee1fbf6763c0d463c377c2e47d8ef2c4f6f.

This originally appeared as a C driver test failure in CDRIVER-3708.