Improve connection pool read performance

[rkistner]

The old implementation worked as follows:

1. If no idle connection is available and the maxReaders haven't been reached, open another connection.
2. Request a read lock on all open connections.
3. The first available read lock is used. Others are released as soon as the lock was available.

This worked okay for a small number of connections, but scaled quite badly when the number of connections is increased. Even with the default maximum of 5 read connections, the number of read locks/s benchmark is reduced from 107k/s to 36k/s on my machine when those connections have been opened. This adds significant overhead when doing many individual select statements.

This is a complete re-implementation of the logic, using a queue of pending operations and available connections instead. This ends up being simpler (less edge cases), and scales much better with the number of connections.

Builds on #37 Now part of this PR:

Fixes #36 (issues when closing the database while queries are still running).

Also improves handling of database initialization errors and other uncaught errors in connection isolates.

[stevensJourney]

Awesome! LGTM