remove internal RPC in the server

[dsschult]

@ddelventhal asked a long time ago why we had modules in separate processes with IPC connecting them. At the time, this was necessary for throughput (keep modules in separate processes because of the GIL). I believe with the advent of futures and asynchronous python programming, this problem can now be completed without the internal RPC and separation of modules.

See https://github.com/dsschult/file_catalog for an example of running tornado + executor pool.

Strategy:

• main process runs webserver
• periodic callback to run queueing on executor
• db runs in executor pool
• schedule and master_updater run off main ioloop
• config and messaging modules go away

Advantages:

• remove a lot of the complexity with RPC and multiprocessing
• yield statements make code easier to follow than callbacks

Disadvantages:

• (none that I can find)

[dsschult]

Changing strategy slightly:

• main process runs webserver
  • fork if not running sqlite
  • this gives us our "parallelism" for db methods
• periodic callback to run queueing on main ioloop
  • only if tornado.process.task_id() == 0
• db methods run off main ioloop
• db queries run in executor thread pool
• schedule and master_updater run on main ioloop
  • only if tornado.process.task_id() == 0
• config and messaging modules go away

[dsschult]

diff stat:

 35 files changed, 2581 insertions(+), 5877 deletions(-)

And that's without touching the tests yet.

[dsschult]

groups branch is now merged:

83 files changed, 7224 insertions(+), 13275 deletions(-)