Support for Prometheus staleness markers

[jmacd]

There are several issues in the Prometheus WG that cannot be addressed without specifying how to treat staleness markers.

Mentioned in open-telemetry/wg-prometheus#59
Mentioned in open-telemetry/wg-prometheus#38

None of these issues actually state that we need a way to represent the NaN value, particularly for Histogram data points which do not always have a semantically meaningful floating point value where a NaN value could go (the sum is not always meaningful, see issues #303, #314).

Prometheus used a specific IEEE 754 NaN value. In #309 we've discussed using a bit for this.