This is the 6.6/6.7 implementation of a master node service to
keep track of the native process memory requirement of each ML
job with an associated native process.

The new ML memory tracker service works when the whole cluster
is upgraded to at least version 6.6.  For mixed version clusters
the old mechanism of established model memory stored on the job
in cluster state is used.  This means that the old (and complex)
code to keep established model memory up to date on the job object
cannot yet be removed.  When this change is forward ported to 7.0
the old way of keeping established model memory updated will be
removed.

…emory

Persistent tasks will recheck allocations when custom metadata
changes, so updating a timestamp in the ML metadata will enable
persistent tasks whose allocation was deferred to allow for a
memory refresh to have another go at selecting an ML node.

Adding this test showed that there was a problem with using
Instant.now() in a cluster state update.  It is possible that
the cluster state update can get applied on more than one master
node if the cluster is unstable, and it's essential that they
both apply exactly the same change.  Therefore I changed the
discriminant to the old cluster state version plus one.  The
actual time is not required.  We just need a field whose change
will kick persistent tasks.

Also added test logging in case it fails again