`getindex(vi::AbstractVarInfo, ::Colon)` docstring is wrong

[penelopeysm]

DynamicPPL.jl/src/abstract_varinfo.jl

Lines 150 to 154 in e4fa7f2

	"""
	getindex(vi::AbstractVarInfo, ::Colon)
	Return the current value(s) of `vn` (`vns`) in `vi` in the support of its (their)
	distribution(s) as a flattened `Vector`.

This is clearly not true, we regularly use vi[:] to access a vector of params as stored internally.

julia> using DynamicPPL, Distributions

julia> @model f() = x ~ Beta()
f (generic function with 2 methods)

julia> m = f(); vi = VarInfo(m); vi2 = link(vi, m);

julia> vi[:]
1-element Vector{Float64}:
 0.4059923090008866

julia> vi2[:]
1-element Vector{Float64}:
 -0.3805580510140428

Note that the default implementation

DynamicPPL.jl/src/abstract_varinfo.jl

Line 160 in e4fa7f2

Base.getindex(vi::AbstractVarInfo, ::Colon) = values_as(vi, Vector)

just calls getindex_internal:

DynamicPPL.jl/src/varinfo.jl

Line 2117 in e4fa7f2

values_as(vi::VarInfo, ::Type{Vector}) = copy(getall(vi))

DynamicPPL.jl/src/varinfo.jl

Lines 685 to 690 in e4fa7f2

	getall(vi::TypedVarInfo) = reduce(vcat, map(getall, vi.metadata))
	function getall(md::Metadata)
	return mapreduce(
	Base.Fix1(getindex_internal, md), vcat, md.vns; init=similar(md.vals, 0)
	)
	end

In theory, the simplest fix to this would be to change the docstring to say that it returns possibly-transformed values.

However, it is technically inconsistent that vi[:] gives internal storage (which may be transformed) whereas vi[@varname(x)] will give you the untransformed values. The correct solution should be to fix getindex(vi, :) to give the untransformed values, and any time we need the transformed values, we should be using getindex_internal(vi, :). This will be pretty annoying, because the 'abuse' of vi[:] is everywhere in the codebase.