connection_pool: proper execution semantics for thread_safe=false

[anarthal]

While async_get_connection respects the usual Asio semantics for pools with thread_safe=false, connection tasks, which are technically children agents of async_run, do not. The proper semantics should be:

• For thread_safe=false:
  • async_run intermediate handlers (e.g. notification timers): usual Asio semantics (as is today)
  • Connection task final handlers: since these affect pool state, usual Asio semantics (i.e. if async_run token has a bound executor, use this; otherwise, use the pool's executor). Currently, we always use the pool's executor. EDIT: these should always use the pool's executor, as it is today.
  • Connection task intermediate handlers (e.g. the ones run by connection::async_connect): if async_run token has a bound executor, use this. Otherwise, use the connection's executor. EDIT: these should always use the pool's executor, as it is today.
• For thread_safe=true:
  • async_run intermediate handlers (e.g. notification timers): use the strand (as is today)
  • Connection task final handlers: use the strand (as is today)
  • Connection task intermediate handlers: there's no need to use the strand here, as these don't access pool state. EDIT: these can cause race conditions on the connections, too, so should use the pool's executor.

These should be documented.

EDIT: following the discussion in Slack, and given that using async_run token's for connection tasks is difficult, the current approach is good, but should be documented.