canonicalize Composite{Any}

[mzgubic]

What do we think about the following:

julia> c = Composite{Any}(;a=1, b=2);

julia> canonicalize(c)
ERROR: ArgumentError: type does not have a definite number of fields
Stacktrace:
 [1] fieldcount(::Any) at ./reflection.jl:725
 [2] #s15#27 at /Users/mzgubic/JuliaEnvs/test/alex/dev/ChainRulesCore/src/differentials/composite.jl:189 [inlined]
 [3] #s15#27(::Any, ::Any, ::Any) at ./none:0
 [4] (::Core.GeneratedFunctionStub)(::Any, ::Vararg{Any,N} where N) at ./boot.jl:543
 [5] canonicalize(::Composite{Any,NamedTuple{(:a, :b),Tuple{Int64,Int64}}}) at /Users/mzgubic/JuliaEnvs/test/alex/dev/ChainRulesCore/src/differentials/composite.jl:166
 [6] top-level scope at REPL[5]:1

I suppose this should never really happen (but it does in higher order AD in Zygote).

Would it make sense to define something like

canonicalize(c::Composite{Any}) = c

since in that case we don't know what the fields are, or do we prefer the error to be thrown?

[oxinabox]

Would this just be for Any or would it be for any abstract type?

I think this might make sense, especially for Any.
Where it is like "I am opting out of ChainRules trying to help me out by knowing about the primal".
It would go well with making +Composite{Any} work with any composite

[oxinabox]

I am not sure we have canonicalize right, but what we do have is kinda a hack to do things with e.g. Zygotes current way.
And this makes that hack more useful.
So I think we should do it.