Consider moving initialstate_distribution to the common section in the documentation

[lassepe]

When implementing the POMDP interface -- even when implementing the generative version -- it seems to me that there is barely anything that one can do with this model unless initialstate_distribution is implemented. E.g. when running a HistoryRecorder, even when using the NothingUpdater the code fails, since the initial belief distribution can not be determined (trying to call initialstate_distribution). So the common workaround seems to be to implement some custom distribution type and then define Base.rand on this type which simply calls initialstate from the generative interface.

Maybe initiastate_distribution should be moved to the "common" section of the POMDPs.jl API documentation? Or maybe one could place a hint somewhere that the above workaround exists. @zsunberg suggested that this workaround could be placed in POMDPModelTools.jl in a generalized form.

Feel free to come up with a better title for this issue.

[zsunberg]

Perhaps there should be

initialbelief(::Updater, ::POMDP)

in the interface, with a default of

initialbelief(u::Updater, m::POMDP) = initialize_belief(u, initialstate_distribution(m))

Then things like particle filters can just use initialstate() directly.

[zsunberg]

Yes, it seems like this should be moved to the "Defining Basic (PO)MDP Properties" section.