Documentation for targeting durations

[RomeshA]

Durations used for transitions may not behave as expected when averaging. In particular, consider the case where half of the population has zero probability of a transition, and the other half has a nonzero probability. Working with probabilities, the overall population probability would be half of the nonzero probability. But the durations can't be averaged, because one of them is infinite. More generally, the conceptual issue is that the zero probability/infinite duration is on the assumption that that portion of the population is never eligible for any other transitions, whereas with probabilities, it's like sampling a different group of people to have zero probability at each timestep (so in the duration conceptualization, there is no mixing between the individuals, whereas with probabilities, there is).

In general, this means that targeting duration parameters may produce unexpected results if there are large differences in the durations across the baseline and program outcome values. This may not always be relevant in cases where the duration parameter is not associated with any transitions, so the resolution is to document this behaviour properly

[Rowanmh]

In general the recommendation would be that where programs target a duration parameter that will be used for transitions, converting that parameter to a probability and targeting the probability should be strongly considered to produce the expected results.

[RomeshA]

This could be an ideal use case for an optional warning as per #351 to be raised if the framework targets a transition parameter that is in duration units