Fix flaky test_timeout

We had an interesting test failure in CI where this test failed twice in
a row: we were expecting a timeout in 0.001 seconds, but both times it
took more than one second!

How can this happen? Well, when we ask a request to timeout after N
seconds, here's what we know:

 * the request will time out, eventually
 * it won't timeout before those N seconds
 * it could timeout a long time after those N seconds, especially on a
   busy CI service

Let's take advantage of those facts by specifying a request-specific
timeout that's longer than the pool-level timeout and making sure that
the timeout was long indeed.

We still test that the request-specific timeout overrides the pool-level
timeout, but will stop failing on busy CI services.

Author: pquentin

test/with_dummyserver/test_connectionpool.py

@@ -93,16 +93,16 @@ def test_timeout(self):
 
         # Request-specific timeouts should raise errors
         with HTTPConnectionPool(
-            self.host, self.port, timeout=LONG_TIMEOUT, retries=False
+            self.host, self.port, timeout=short_timeout, retries=False
         ) as pool:
             wait_for_socket(ready_event)
             now = time.time()
             with pytest.raises(ReadTimeoutError):
-                pool.request("GET", "/", timeout=short_timeout)
+                pool.request("GET", "/", timeout=LONG_TIMEOUT)
             delta = time.time() - now
 
-            message = "timeout was pool-level LONG_TIMEOUT rather than request-level SHORT_TIMEOUT"
-            assert delta < LONG_TIMEOUT, message
+            message = "timeout was pool-level SHORT_TIMEOUT rather than request-level LONG_TIMEOUT"
+            assert delta >= LONG_TIMEOUT, message
             block_event.set()  # Release request
 
             # Timeout passed directly to request should raise a request timeout