Computing likelihood for decapitated ARGs.

[GertjanBisschop]

As discussed in #2232, we could easily extent the current log_arg_likelihood such that we also return the correct likelihood when the ARG simulation was stopped early. This means we need to distinguish between roots that are unary across the ARG versus those that are only unary across parts of the ARG.
I guess the easiest solution would be if those nodes had their own flag/NodeType instead of flag=0.