Store state node IDs in parameters

[visr]

So we can remove id_from_state_index.

[visr]

These are scraps from #2077. Thinking about this, and having to deal a bit more with huge stacktraces from the large ComponentArrays types, made me consider another approach. Better in a separate PR, but dumping thoughts here. We could store the UnitRange{Int64} of each state component in the parameters, and use view(u, p.is_pump) instead of u.pump. Storing the resulting SubArray like in StateVector from #2077 can maybe also be done, but is more complex because you'd have to support Duals and GradientTracer as well as Float64.

I'm hoping this can also help with latency (#1942). Since ComponentArrays puts the full Axes into its type signature, it will specialize differently on each model. Because if the state length or component lengths differ, it needs to compile separate specializations.

# uses MethodAnalysis, f(x) can be anything
julia> specializations = methodinstances(f)
2-element Vector{Core.MethodInstance}:
 MethodInstance for f(::ComponentVector{Int64, Vector{Int64}, Tuple{Axis{(a = ViewAxis(1:3, Shaped1DAxis((3,))), b = ViewAxis(4:5, Shaped1DAxis((2,))))}}})
 MethodInstance for f(::ComponentVector{Int64, Vector{Int64}, Tuple{Axis{(a = ViewAxis(1:2, Shaped1DAxis((2,))), b = ViewAxis(3:5, Shaped1DAxis((3,))))}}})

Here the two ComponentVectors only differ slightly:

julia> using ComponentArrays
julia> a = ComponentVector(; a = [1,2,3], b = [4,5])
julia> b = ComponentVector(; a = [1,2], b = [3,4,5])
julia> typeof(a)==typeof(b)
false

Many methods that we call are likely to already be precompiled for simple Vector{Float64}, but none will be for our ComponentVector state (or StateVector).