Pool should spawn a task when connecting in `acquire()`

[abonander]

This way, if the acquire() call is cancelled or times out, the connection can still go through.

The connection task should have its own separate timeout, possibly surfaced as PoolOptions::connect_timeout().

This will likely fix:

• could not receive data from client: Connection reset by peer #3314 #3315
• Pool creates many connections when under high contention #3132
• PoolInner::acquire() does not try the idle queue after a transient connection failure #2848 (PoolInner::acquire() does not try the idle queue after a transient connection failure #2848 (comment))