Remove r2d2_sqlite dependency

[link2xt]

No description provided.

[link2xt]

The whole r2r2 connection pool is an overkill for DC usage. Not being able to reliably close the database due to a long-standing issue sfackler/r2d2#99 causes problems like #4049 where the test migrates the database before it is closed and cannot determine when it will be closed.

[flub]

It'd be nice to get rid of r2d2 entirely someday. We really don't get any benefit out of multiple connections and may as well have a single tokio task that handles all the Db requests communicating over some channels. But that's just dreaming of the future.

[flub]

Unfortunate that we need to keep this in memory the whole time.

[link2xt]

This is also because of r2d2 abstraction, it expects that connections may fail because underlying network connections fail, and have to be reestablished. Which is not the case with SQLite.

[link2xt]

Fixed in #4053.

[link2xt]

It'd be nice to get rid of r2d2 entirely someday. We really don't get any benefit out of multiple connections and may as well have a single tokio task that handles all the Db requests communicating over some channels. But that's just dreaming of the future.

This is what sqlx does, it creates "connection workers" which receive commands and send back results over channels. We previously dropped sqlx because it accessed the same connection from different threads and had performance problems as a result. This problem is likely solved now, but I would not try to reintroduce sqlx again for now, first want to solve r2d2 problems the simplest way, likely by putting a single connection behind RwLock and just using it.