MesosPathDetector double-counts checkpoint roots

This manifests itself in the observer ballooning in resource usage. The TaskDetector finds two copies of every task because of the 'latest' symlink being a duplicate of the actual directory, so we get race conditions in on_active/on_finished inside src/main/python/apache/thermos/observer/task_observer.py that cause it to hold onto TaskResourceMonitors which tend to be pretty expensive to run.