Improve metrics collection strategy

[krancour]

#15 and #16 will both require us to alter our approach to serving and aggregating metrics, since it is clear that request count is an inadequate metric upon which to make scaling decisions if we are to support protocols other than HTTP (e.g. HTTPS, HTTP2, gRPC).

The new approach may "simply" (it's not actually simple) involve counting active TCP connections over a period of time.

I'd like to suggest avoiding a dependency on Prometheus for as long as it can be avoided, as it raises the barrier to entry for using Osiris. In the near term, we can investigate options for metrics-collecting sidecar proxies to keep time series data in memory, with older data evicted as is become irrelevant.

[krancour]

This was fixed by #27.