mds: optimize acquiring locks

mds: optimize acquiring locks

There are several changes:
 - use single LockOp vector to pass all locks (rdlock, wrlock, xlock)
   to Locker::acquire_locks(). this avoids memory allocation overhead
   of using multimple set
 - change MutationImpl::locks to LockOp set, use it to track both locks
   and locks' types.
 - use empalce_hint to optimize insertion of lock to MutationImpl::locks
 - use iterator to optimize removal of lock from MutationImpl::locks

Signed-off-by: "Yan, Zheng" <zyan@redhat.com>