Rewrite the async job system

[sandhose]

We're currently using apalis for async jobs, which… is probably not mature enough.
A few issues with it:

• the crates are a bit messy
• they are doing breaking changes all the time. See Upgrade apalis #2784
• the database structure is… weird
• the jobs don't retry correctly
• the "workers" crash when it looses the connection and don't restart
• sharing state in the job is not type safe

Requirements:

• have jobs in the database
• use triggers to NOTIFY workers for new jobs (optional?)
• lock rows with PG's FOR UPDATE SKIP LOCKED
• pass the connection which locked the task row to the handler
• have a way to do cron-like schedules for maintenance tasks

Related:

• Losing the connection to Postgres stops the background jobs #2625
• Ensure background jobs are reliably retrying #1490

[matrixbot]

For your information, this issue has been copied over to the Element fork of matrix-authentication-service: element-hq/matrix-authentication-service#2785