Previously on this branch, it would attempt to create a superuser and
initialize the replica set after the local host exemption had passed
(once a mongod is started with a keyfile and the replicationSetName
 config file option or --replSet command line argument, you can't create
 users using the localhost exemption anymore ).

This necesarily complicates the logic and requires us to track if we're
initializing a replica set or not (with facts) and to template out
/etc/mongod.conf twice in the initialization case.

Further modifications to the replica set can be run against all the
members and will install mongo on new machines while updating the
replica set on the old machines.  With a further commit on this branch,
we can avoid needlessly restarting mongo on all machines in the cluster.

When we run this on an existing cluster, this stops and starts mongo for
all machines it is run against.  We can't rely on the notify: since that
happens at the end of the play and during initial cluster setup, we need
to restart and pick up the changed config.

Try to stick sensible tags on things

Ensure mongo restarts on the cluster key file change

Add a few playbook notes about common operations

Remove check for mongo 2.4 since this is the mongo_3_0 role
This was probably carried over from the mongo 2 role. I see no reason
why mongo 2.x and mongo 3 couldn't coexist on the same machine.

Remove old hugepages init script check
This was probably added while we were still iterating on our mongo 3
deployment, but it should no longer be necessary.

Clean up ansible syntax

Don't move the old mongo data dirs
This actually skips for edX because we provision machines that already
have {{mongo_data_dir}} mounted on an external disk.  However, for
non-edX use, this could fail if you turn on WiredTiger since it will
move mmapv1 files into the mongo_data_dir and then mongo will fail to
start because it has been told to use WiredTiger.

Don't make this a serial play
We usually run this on 3 machines, so serial: 3 was equivalent to
ansible's default of "run everything in parallel" but this causes
problems if you ever run it on 4 like we do for prod envs.
In addition, this prevents run_once from working properly, and there are
a number of things we only one to do on one machine (like creating a
superuser).

We can't use the 3.0.7 tools (mongodump/mongorestore) because of bugs in
the restore on gridfs and even draft mongo modulestore.
https://jira.mongodb.org/browse/TOOLS-939

This does not affect the mongo shell version

The * was rotating already rotated logs
OPS-1009

It turns it into key/value arguments to the ansible module, causing
hilarious failures.

Explanation: if mongo is set with systemLog.logRotate='reopen', when it
receives a SIGURS1, it'll close and reopen its log file handle.

Changes were originally introduced in 2c53e32c, PR #2217

Changes were originally introduced in 02cda120, PR #2214