connect_timeout is applied per resolved address, not per endpoint

`connect_stream` applies `connect_timeout` **per resolved address**, not per endpoint, contradicting the doc comment ("TCP connect timeout per endpoint attempt"). A hostname resolving to two addresses multiplies every bound: measured 2.0011s for two blackhole addresses at a 1s timeout.

Fix: give `connect_stream` a single total deadline across all resolved addresses, so the timeout bounds the endpoint attempt as documented.

Start by reading connect_stream and its connect_timeout doc comment, then reproduce the two-address case described in the issue to confirm the current per-address timing. Done means one total deadline is shared across all resolved addresses so the endpoint attempt stays within the documented timeout.

Written by the indexing model from the issue text.