Support multiple file writers writing at once.

Background

TensorBoard assigns one DirectoryWatcher per run. The DirectoryWatcher watches for new events being written to the directory and loads it into TensorBoard's multiplexer. Today, the DirectoryWatcher iterates through events files (for a run) in lexicographic order. After it has finished reading from an events file, it never reads from it again (and moves on to a subsequent events file).

This behavior means that TensorBoard does not support multiple file writers writing to the same run directory: TensorBoard would move on to a different events file B once it finishes reading events file A and would never go back to reading A despite how A is updated.

In turn, that problem blocks TensorFlow+TensorBoard's transition towards migrating Estimator, tflearn hooks, and other high-level TensorFlow constructs to using the tf.contrib.summary-style summaries. tf.contrib.summary-style summaries are written via a type of SummaryWriter different from the one that writes traditional-style (tf.summary.*) summaries, so users that intermingle the summaries will necessarily have to concurrently use 2 summary writers. Many users thoughout google3 intermingle the summaries, and the transition must happen soon.

Start with tensorboard/backend/event_processing/directory_watcher.py, especially the lexicographic event-file iteration described in the issue. Trace how DirectoryWatcher tracks completed files and verify behavior when two writers update the same run directory; done means events from both files continue reaching TensorBoard's multiplexer.