make rand_tangent of struct with no perturbable fields return DoesNotExist

[oxinabox]

Closes #149
This is my preferred alternative to JuliaDiff/ChainRulesTestUtils.jl#142 (comment)

Note that it is not the case that if all my fields do not exist then i do not exist,
but it is the case that if all my fields are not perturbable then I am not pertablable.
Which is what DoesNotExist actually means, and we should rename it.

[oxinabox]

The outer if here that gives us the NO_FIELDS would be able to be removed if we replaced NO_FIELDS with DoesNotExist().
Since then we would be in the case that all the fields (of which there are none) subtype DoesNotExist.
Vacuous truth.

JuliaDiff/ChainRulesCore.jl#175 (comment)

[mzgubic]

Also closes #153

[mzgubic]

Actually, it won't close any of the issues, because we still have two copies of rand_tangent (one here and one in ChainRulesTestUtils).

[oxinabox]

Also closes #153

It won't close that

julia> rand_tangent(Random.GLOBAL_RNG, Colon())
ChainRulesCore.Zero()

It would close that if we also did
JuliaDiff/ChainRulesCore.jl#175 (comment)
which i think we should (which would also simplify the code in this)

Actually, it won't close any of the issues, because we still have two copies of rand_tangent (one here and one in ChainRulesTestUtils).

[mzgubic]

Oh yeah that's right, my bad. I am happy to move the code over and make it work. We discussed on slack a while back.

[oxinabox]

Thanks.
Please do so.
Either location is better than both

[mzgubic]

I think we can merge this now since JuliaDiff/ChainRulesTestUtils.jl#147 is in

[mzgubic]

Well, actually, needs a version bump, not sure why it does not show up as conflict.