Run integration tests with an in-memory sqlite database instead of a hand-coded fake database

[maxbrunsfeld]

As the collab server starts to store more and more state in its database (as opposed to in memory), it will become more and more inconvenient to maintain our "Fake Db", which we currently use for integration tests, in order to avoid the slowness of Postgres.

This PR replaces the Db trait (which previously had two implementations: fake and Postgres) with a generic struct that works with both Postgres and Sqlite. Some database methods are not implemented for Sqlite (since they aren't run during integration tests). And some methods are implemented slightly differently for Sqlite than for Postgres, because of differences in the SQL dialect. But most of our queries work identically against Postgres and Sqlite.

Like before, we have tests for the Db itself which assert that the Sqlite and Postgres implementations behave the same.

Note - This PR also removes all the remaining code for chat and for project activity tracking. Those features were both no longer used, and were complicating our database work, so we decided to go ahead and 🔥 them.

Note - This updates our sqlx dependency to use a fork, because of this change that I needed to make to sqlx's dependency on libsqlite-sys.