Caution with unused variables in SimpleVarInfo

[yebai]

The new VarInfo #309 introduced many nice new features and significant performance improvements. However, it needs to be used with caution at the moment. For example, we can easily introduce spurious variables into SimpleVarInfo which does not influence logdensity evaluation, but would subtly change HMC behaviour (for good or bad). It is because the spurious variables (i.e. x[3:4]) would introduce additional momentum variables. Since these variables do not have a prior on them, the sampling results would become slightly tricky to analyse.

julia> using DynamicPPL, StableRNGs, Distributions

julia> @model function demo()
           m ~ Normal()
           x = Vector{Float64}(undef, 2)
           for i in eachindex(x)
               x[i] ~ Normal()
           end
           return x
       end

julia> _, vi = DynamicPPL.evaluate!!(m, SimpleVarInfo((x = ones(2), )), ctx);

julia> _, vi = DynamicPPL.evaluate!!(m, SimpleVarInfo((x = ones(4), )), ctx); # Note the spurious x variables x[3:4]

julia> vi
SimpleVarInfo((x = [-1.6438528680920432, 0.7944393211208289, 1.0, 1.0], m = 1.2973461452176338), -5.265062153308225)