Default float type to float(Real), not Real

[penelopeysm]

Closes #684

The issue seen in #684 results from this series of calls:

DynamicPPL.jl/src/model.jl

Lines 1190 to 1200 in cdd3407

	function Distributions.loglikelihood(model::Model, chain::AbstractMCMC.AbstractChains)
	var_info = VarInfo(model) # extract variables info from the model
	map(Iterators.product(1:size(chain, 1), 1:size(chain, 3))) do (iteration_idx, chain_idx)
	argvals_dict = OrderedDict(
	vn_parent =>
	values_from_chain(var_info, vn_parent, chain, chain_idx, iteration_idx) for
	vn_parent in keys(var_info)
	)
	loglikelihood(model, argvals_dict)
	end
	end

Here, argvals_dict is an OrderedDict, and when likelihood() is called on L1198 it gets promoted to a SimpleVarInfo where the type of logp is float_type_with_fallback(infer_nested_eltype(typeof(argvals_dict))):

DynamicPPL.jl/src/simple_varinfo.jl

Lines 214 to 222 in cdd3407

	function SimpleVarInfo(θ::Union{<:NamedTuple,<:AbstractDict})
	return if isempty(θ)
	# Can't infer from values, so we just use default.
	SimpleVarInfo{SIMPLEVARINFO_DEFAULT_ELTYPE}(θ)
	else
	# Infer from `values`.
	SimpleVarInfo{float_type_with_fallback(infer_nested_eltype(typeof(θ)))}(θ)
	end
	end

In this case, infer_nested_eltype(typeof(argvals_dict)) is Any and thus float_type_with_fallback(Any) gives Real. And then when resetlogp!! is called, it sets logp to zero(Real), which ends up being an Int64 (on my system).

This PR proposes to make float_type_with_fallback(Any) return float(Real) instead, which is Float64.

[penelopeysm]

I'm proposing to release a bugfix for 0.28 first, since we're still working on a version of Turing that's compatible with 0.29 (TuringLang/Turing.jl#2341). But if we merge this, I can ofc port this to the master branch too.

[mhauru]

Makes sense to me

[mhauru]

Even better with a test. :) Thanks @penelopeysm