Restructure the WSGI app for easier direct use

[nfelt]

This issue tracks refactoring the main TensorBoard WSGI app to make it easier to use the app as a standalone WSGI app (rather than only via the tensorboard CLI).

The essential stream of work is the "core" stream below, but it leaves behind a lot of cleanup that we should at least partially address as well.

CORE

• Migrate callers of TensorBoardWSGIApp users to TensorBoardWSGI + explicit Reload() - done in Refactor away TensorBoardWSGIApp in its current form #2575
• Reorganize the logic in standard_tb_wsgi() so that first half creates the multiplexer + launches reloading, second half creates the TBContext and instantiates plugins - done in Refactor away TensorBoardWSGIApp in its current form #2575
• Factor out second half of standard_tb_wsgi() as a reincarnated TensorBoardWSGIApp entry point that accepts data_provider (and multiplexer as the legacy option) - done in Reincarnate TensorBoardWSGIApp as new DataProvider-friendly entry point #2576
• Encapsulate standard_tb_wsgi() multiplexer creation and reloading logic into new LocalDataIngester construct - done in refactor: slice the two-headed application.py monster in half #3643
• Move multiplexer-related bits of application.py into a separate file - done in refactor: slice the two-headed application.py monster in half #3643
• Refactor program.py to create LocalDataIngester manually, and get rid of standard_tensorboard_wsgi() - done in refactor: slice the two-headed application.py monster in half #3643

CLEANUP: WSGI

• Create dispatch middleware for mapping WSGI apps to paths
  • some example implementations: werkzeug, cheroot
• Change path_prefix to be application of dispatch middleware to TensorBoardWSGIApp, rather than built-in feature - partly done in application: extract path prefix logic to middleware #2733
• Make each plugin itself a WSGI app, so that their backends can be tested entirely independent of the main TB app, and remove calls to TensorBoardWSGI
  • could do this by implementing __call__ in BasePlugin using dispatch middleware plus get_plugin_apps()?
  • could change get_plugin_apps() to memoized default implementation that searches class for attributes carrying a special tag, and use a method decorator to apply this tag to the individual routes (e.g. @tb_route("/tags") def tags_route(self, request): ...)
• Refactor away TensorBoardWSGI inner surface (move plugin route aggregation logic into a helper that we can test directly)
• Change TensorBoardWSGIApp.__call__() to just use dispatch middleware logic

CLEANUP: CORE PLUGIN

• Move CorePlugin flag definitions into other places as described in Eliminate passing of untyped CLI flags across API boundaries #2801
• Move CorePlugin routes onto TensorBoardWSGIApp itself

[nfelt]

Factored out the CorePlugin flags cleanup into #2801.

[nfelt]

Update: several more pieces done in #3643.