For simulating overloaded backends it would be useful to be able to have a means
to increase the injected delay as the number of requests behing handled grows.

There's the downstream_rq_active gauge that looks like it could be used as an
input to this. One way to achieve this would be to add configuration settings that
control this computation to allow one to have latency computed based on the delta
between the current and target value of a gauge [1]

The final delay would then look like (target-value-of-the-gauge - current-value-of-the-gauge) * delay_step_us.
To trigger the fault filter, the test server extension could synthesize an x-envoy-fault-delay-request: <computed value> header.

[1] example

          - name: test-server
            config:
              stats-based-delay: 
                  - gauge: "downstream_rq_active":
                     target-gauge-value: 100
                     delay-step-us: 1000

It's tempting to generalize this more by integrating the CEL expression engine, but it looks like that is an order
of magnitude more work (I may be wrong about that, but that is my gut feeling after skimming some code).