Standardise use of OffsetVectors

[THargreaves]

This PR modifies all GeneralisedFilters/SSMProblems code to use use an OffsetArray for storing the xs and zs. Docstrings are updated to match.

This cleans up a few bits of the codebase and solves #49.

[FredericWantiez]

Might be missing something, but why not something like this:

    xs = Vector{T_dyn}(undef, T)
    ys = Vector{T_obs}(undef, T)

    xs[1] = simulate(rng, model.dyn; kwargs...)
    for t in 1:T
        ys[t] = simulate(rng, model.obs, t, xs[t]; kwargs...)
        if t < T
            xs[t+1] = simulate(rng, model.dyn, t, xs[t]; kwargs...)
        end
    end

[THargreaves]

It's just how we've happened to define the SSMProblems interface. I recall that Hong was keen to have a separation between the initial state x0 and the other states that have associated observations.

Following from that convention, I feel that the offset vector is the cleanest way to achieve it.