Rewrite core_plugin_test.py to use DB import mode instead of tf.contrib DB writer

[nfelt]

This removes the tf.contrib.summary.create_db_writer() dependency from core_plugin_test.py, which partially addresses #1718 and #1705, and reverts the applicable part of #3143 (a temporary fix).

I've preserved most of the test coverage of the underlying DB-reading code by changing the test data population strategy to instead just use the DbImportMultiplexer logic that already exists within TensorBoard. That logic doesn't set started_time so we lose some coverage in that the test that specifically tests that behavior is now logdir-only, because it didn't seem worth the trouble to get it to keep working.

I also refactored the rest of the tests a little bit to remove cruft and reorganize the test cases a little more logically. Ultimately a lot of this should be cleaned up further as part of #2573. (And ultimately we should either drop the DB reading logic entirely or encapsulate it inside a DataProvider, but that's a battle for another day.)

[stephanwlee]

Before merge, would you kindly remind me how CorePluginDbMode differs from CorePluginDbImportMode?

[nfelt]

It's basically the branch here:

DB import mode:

tensorboard/tensorboard/backend/application.py

Line 131 in 94faac5

if flags.db_import:

DB read-only mode:

tensorboard/tensorboard/backend/application.py

Line 148 in 94faac5

# DB read-only mode, never load event logs.

Both modes exercise the SQL-based read path within the plugins, which is what we actually want to test. The difference is that DB read-only mode expects you to pass an existing SQLite DB file with the --db flag (and the old test was using the create_db_writer() code to create such a file), whereas the DB import mode takes a logdir and automatically imports the logdir event files into a temporary DB (without using create_db_writer() but instead implementing mostly equivalent behavior).