gh-157174: Close the socket when HTTPConnection.connect() fails to set TCP_NODELAY

[iamsharduld]

HTTPConnection.connect() assigns the new socket to self.sock and then sets TCP_NODELAY on it. When that setsockopt() fails with anything other than ENOPROTOOPT, the error propagates while self.sock still holds the open socket, and a caller that drops the connection leaks it. _tunnel() handles its own connect-time failure by calling self.close() before raising; this does the same for TCP_NODELAY.

It is what is behind the ResourceWarning from test_ssl.test_https_client_non_tls_response_ignored on macOS CI (details in the issue): on macOS, setsockopt(TCP_NODELAY) raises EINVAL once the peer has reset the connection, that test's server resets on purpose, and on a slow runner the reset wins the race.

Verified on macOS 26:

• A harness with a server that accepts and resets at once, and a client whose _create_connection pauses 2 ms before returning (standing in for the client thread being preempted): 150 of 150 attempts raise EINVAL from setsockopt(); without this change all 150 sockets are reported unclosed, with it none.
• The new test_connect_tcp_nodelay_error_closes_socket fails before the change (conn.sock is not None) and passes after; test_connect_tcp_nodelay_unsupported checks that ENOPROTOOPT still leaves the connection usable.
• test_httplib, test_urllib2, test_urllib2_localnet and the test_ssl test in question pass.

• Issue: http.client.HTTPConnection.connect() leaks the socket when setting TCP_NODELAY fails (EINVAL on macOS after a reset) #157174