Own a direct connection by the process that opened it

[benoitc]

A connection opened without a pool was started under hackney_conn_sup, so the supervisor was its owner. When the caller was killed, nothing closed it, and with idle_timeout at infinity the socket stayed open until the server closed it. Found in barrel, where a query worker killed at its deadline left its connection open.

The start sites now pass owner => self(), since they run in the caller:

• hackney:connect_direct/4: request/5 and connect/* with {pool, false}, streamed request bodies, async responses, plain HTTP proxies, and HTTP/2 and HTTP/3 without a pool
• hackney:start_conn_with_socket_internal/5: CONNECT and SOCKS5 tunnels
• hackney_h2_stream:establish/2: h2_open/*, so the connection closes even when the stream process is killed before terminate/3

With stream_to set to another process, that process owns the connection. This was already the case (#646) and is now tested. The pool path is unchanged, and set_owner/2,3 still moves ownership.

hackney_direct_owner_tests covers each entry point: kill the opener, then wait for the conn's 'DOWN' and the server-side socket close. No sleeps. 10 of its 12 tests fail on 4.8.1. The stream_to and set_owner tests pass there too, as expected. unpooled_conn_not_shared now asserts the uploader is monitored once, as owner.

Known limit: an HTTP/1.1 conn waiting for a response is blocked in a passive recv, so it sees its owner's death only when the read returns, at the latest after recv_timeout.