Nested PrefixContext behaviour is not fully defined

[penelopeysm]

# Setup
using DynamicPPL
p1 = PrefixContext{:a}(DefaultContext())
p2 = PrefixContext{:a}(PrefixContext{:b}(DefaultContext()))
p3 = PrefixContext{:a}(SamplingContext(PrefixContext{:b}(DefaultContext())))
p4 = SamplingContext(PrefixContext{:a}(DefaultContext()))

julia> prefix(p1, @varname(x))
a.x

julia> prefix(p2, @varname(x))
b.a.x

julia> prefix(p3, @varname(x)) 
a.x

julia> prefix(p4, @varname(x))
ERROR: MethodError: no method matching prefix(::SamplingContext{SampleFromPrior, PrefixContext{…}, Random.TaskLocalRNG}, ::VarName{:x, typeof(identity)})
The function `prefix` exists, but no method is defined for this combination of argument types.

Several points:

1. prefix(p1, @varname(x)) is all good.

2. It seems to me that it would be more natural for prefix(p2, @varname(x)) to return a.b.x as right now this means that nested submodels have the prefixing occurring in the less intuitive way:

using DynamicPPL, Distributions

@model function subsubmodel()
  x ~ Normal(0,1)
end

@model function submodel()
  x ~ to_submodel(prefix(subsubmodel(), :c), false)
  return x
end

@model function parentmodel()
  x1 ~ to_submodel(prefix(submodel(), :a), false)
  x2 ~ to_submodel(prefix(submodel(), :b), false)
end

keys(VarInfo(parentmodel()))
# 2-element Vector{VarName{sym, typeof(identity)} where sym}:
#  c.a.x
#  c.b.x

3. The intuitive behaviour for prefix(p3, @varname(x)) would be to prefix with both a and b

4. prefix(p4, @varname(x)) should traverse the context tree accordingly. This is related to to_submodel does not correctly handle manually specified prefixes when sampling #785, where the underlying problem is that the model is being evaluated with a DebugContext(SamplingContext(PrefixContext...))) but the varnames aren't prefixed.