Add g3_param_project() for both random devates & projections

[lentinj]

Right, I think this is ready now. The final commit is the interesting one.

The most interesting bit to look at is the documentation, which shows how to use it

But essentially you fill a gap for a formula with g3_param_project() and you get per-step/per-year parameters, a likelihood component & projections based around the same parameters the likelhood component uses:

    g3a_naturalmortality(st, g3a_naturalmortality_exp(
        g3_param_project("Mdn", g3_param_project_dnorm(
          mean_f = g3_parameterized("Mdn_mean", value = 0.1, optimise = FALSE),
          stddev_f = g3_parameterized("Mdn_sd", value = 0.001, optimise = FALSE) )))),
    NULL )

The main thing that's irritating me is there's no by_stock option. I've not thought of a non-painful way of doing it, so not added it for now, but it's an obvious thing to want.

We should probably have parametrised defaults for mean/stddev as above. Naming these is also marginally tricky, but not the end of the world.

@bthe any thoughts?

Fixes #181

[bthe]

Yes, this sounds pretty good. There are two scenarios I don't see immediately how we can deal with though:

• Should there be a generic g3_param_project_custom where the user could supply its own likelihood function?
• How would you deal with correlated time-series? In this scenario we could consider $L_\infty$ and $\kappa$ in Von B to be timevarying but from a joint multi-variate normal distribution.

[lentinj]

Should there be a generic g3_param_project_custom where the user could supply its own likelihood function?

You could supply your own list with the pair of formulas, nll & project easily enough, or replace the formula returned by g3_param_project_dnorm. Depends on what your custom likelihood is hoping to do. It'd be a lot of boilerplate to just replace dnorm/rnorm with another pair of functions, but I think if that was the case it'd be easier to add all the variants to gadget3.

[lentinj]

How would you deal with correlated time-series?

We'd need a set of helpers corresponding to TMB's UNSTRUCTURED_CORR / VECSCALE for specifying the nature of the correlation matrix, but that's easy enough. Simulations could be wedged into another native function.

From an interface point of view, the more interesting question is we have a similar problem to M, we have an array of dimensions c(time, c('Linf', K)), but want to refer to the individual parts in different places. We also don't want to have to define this array twice. This is worth having a decent answer to before merging the above.