implement continuation methods for profiles

[bbolker]

as far as I can tell, profiling is using the same starting values as the original call. This is inefficient and particularly bad when the likelihood surface is sufficiently weird that there are multiple modes for focal parameter values away from the MLE. Should implement

• continuation method (i.e. use estimates from previous focal parameter fit)
• linear extrapolation method

[bbolker]

This is at least partly done in 1bb009d (at least the "naive" method), but a linear-extrapolation method might be even better. (Also, not enabled by default yet.)