Improve error message when DifferentiationInterface isn't loaded

[penelopeysm]

Since #39 was merged, it's now possible to run e.g. ADgradient(ADTypes.AutoMooncake(),...) using DifferentiationInterface. 🎉

This is documented here:

LogDensityProblemsAD.jl/ext/LogDensityProblemsADADTypesExt.jl

Lines 25 to 29 in e3401f2

	Passing `x` as a keyword argument means that the gradient operator will be "prepared" for the specific type and size of the array `x`. This can speed up further evaluations on similar inputs, but will likely cause errors if the new inputs have a different type or size. With `AutoReverseDiff`, it can also yield incorrect results if the logdensity contains value-dependent control flow.
	If you want to use another backend from [ADTypes.jl](https://github.com/SciML/ADTypes.jl) which is not in the list above, you need to load [DifferentiationInterface.jl](https://github.com/gdalle/DifferentiationInterface.jl) first.
	"""
	LogDensityProblemsAD.ADgradient(::ADTypes.AbstractADType, ℓ)

but I wonder if one could define a catch-all method, mirroring the one here

LogDensityProblemsAD.jl/src/LogDensityProblemsAD.jl

Lines 66 to 69 in e3401f2

	function ADgradient(v::Val{kind}, P; kwargs...) where kind
	@info "Don't know how to AD with $(kind), consider `import $(kind)` if there is such a package."
	throw(MethodError(ADgradient, (v, P)))
	end

which could mention that if you're expecting your AD backend to work via DI, then try importing DI?

[penelopeysm]

Pfft, I was using :Mooncake instead of AutoMooncake(). 😓 Let me edit to make this better. Edited.