Support master/replica databases

[jace]

For performance reasons, apps like Hasjob in production should have separate master and replica databases, running on different nodes:

1. All writes happen on the master database
2. All reads happen on the replica database
   a. …unless we just wrote to master, in which case read from master

Here are a couple of gists showing how this can be achieved transparently with Flask-SQLAlchemy:

• https://gist.github.com/trustrachel/6828122
• https://gist.github.com/adhorn/b84dc47175259992d406

We'll need a variation of the second gist to handle scenario 2a: record a dirty flag in application context (_app_ctx_stack.top) and send all queries to master for as long as the flag remains. We also need to debug and resolve the problem with self._flushing, as reported in the various comments on the first gist.

Coaster provides a copy of Flask-SQLAlchemy's SQLAlchemy object in the coaster.db module (not the db instance that is also provided). Our version of the subclass could reside here and be available for importing to apps.