This makes the request duration histograms consistent across both
metrics and only records it when the request was successful.

It also makes sure that the way the error was received in the
middleware does not matter anymore. Before this change, an error that
bubbled up towards the middleware would count towards a groups error
budget, but not towards the service availability. With this, both metrics
will not include durations of requests that resulted in responses with
a 5xx status code.

If a request failed, it's not very important for users how fast it
failed.

For our metrics it could also skew results: A very fast 500 would have
less impact on a service's availability or a groups budget spend than
a slow one. While they should weigh the same. Similarly, a slow
failure should not count double towards availability.

This results in the following scoring for requests:

|         | Fast | Slow |
|---------|------|------|
| Success | 2/2  | 1/2  |
| Error   | 0/1  | 0/1  |