Remove all uses of asyncio.shield() in our codebase

[achimnol]

It may make dangling but running tasks when shutting down, because it only cancels the outer await but keeps the inner task intact. asyncio.shield() is useful to protect tasks that are "not well written to react on cancellation" but requires a strong assumption that all shielded tasks should terminate within a finite, short period of time.

Historically aiopg did not handle cancellation of transactions during entering of them (aio-libs/aiopg#332), and we have experienced exhaustion of database connections due to accumulation of unreleased transactions and connections when there are cloud users from unstable WiFi (aiohttp cancels request handler tasks when the connection is interrupted). This is where we started to use asyncio.shield() for some database-critical operations, such as AgentRegistry.enqueue_session() called from API handlers. Now we don't use aiopg with SQLAlchemy but asyncpg, which handles this better.

Let's remove uses of asyncio.shield() and replace it with aiotools.PersistentTaskGroup if necessary.
Let's make separate PRs for each module code update.

I expect that this would reduce the possibility of getting stuck while shutting down services.

[achimnol]

For the asyncio.shield() usescases of short-lived database transaction wrappers in the api.stream module, let's wrap the ptaskgroup.create_task() with asyncio.shield() and apply the shutdown timeout option of ptaskgroup (though achimnol/aiotools#33 must be implemented first) so that database transactions always complete in common scenarios. For now, just wrap them with a ptaskgroup (which is stored in the app-local context).