Extend program.TensorBoard API to allow other servers and absl flags

[nfelt]

This PR extends the program.TensorBoard API introduced by #1240 so that it's compatible with internal use of TensorBoard. It makes the following changes:

• program.TensorBoard() now accepts a server_class parameter specifying a subclass of the new TensorBoardServer abstract class, to customize the serving logic for TensorBoard's WSGI app
• program.TensorBoard() will now use the new absl flags argparse integration if available so that absl flags get parsed together with the regular argparse ones, and any absl flags from the main module are exposed via TBContext.flags
• program.TensorBoard() now accepts plugins as either TBPlugin classes (as it was before Migrate TensorBoard to argparse #1240) or TBLoader plugin loader instances, and constructs BasicLoaders automatically for the TBPlugin classes. This saves some boilerplate and makes it easy to do targeted removal of specific plugins from the default set if desired because they aren't wrapped in an undistinguishable BasicLoader.

The corresponding internal changes are previewed in cl/211881012. Together these get the internal sync unblocked.