add error message for integers that represent reals

[mzgubic]

e.g. example by @SomTambe

julia> f(x) = 8x + 10*sin(x)
f (generic function with 1 method)
julia> function ChainRulesCore.rrule(::typeof(f), x)
           y = f(x)
           function ∇f(Δ)
               ∇ = 8 + 10*cos(x)
               @show ∇ Δ
               return ChainRulesCore.NoTangent(), Δ * ∇
           end
           return y, ∇f
       end
julia> test_rrule(f, 2);
∇ = 3.838531634528576
Δ = 2.31
∇ = 3.838531634528576
Δ = 2.31
test_rrule: f on Int64: Test Failed at /home/somvt/.julia/packages/ChainRulesTestUtils/AX7fv/src/testers.jl:278
  Expression: ad_cotangent isa NoTangent
   Evaluated: 8.86700807576101 isa NoTangent
...
Test Summary:          | Pass  Fail  Total
test_rrule: f on Int64 |    4     1      5
ERROR: Some tests did not pass: 4 passed, 1 failed, 0 errored, 0 broken.

the error message is confusing, we could special case it for actual::AbstractFloat, expected::NoTangent as suggested by @oxinabox

[SomTambe]

thanks for the issue!

[oxinabox]

Similar there are some other cases like AbstractVector{<:Interger} that really represent AbstractVector{<:AbstractFloat}.

We can check the primal type and see if it is one we automatically generated a NoTangent() from and then tell the user that if the are using this type then we can not be sure if they mean to use it as a non-differentiabled quantity like an index (our default assumption) or if they are using it to represent a continuous differentiable quantity.

For the particular case of integers we mike want to suggest rather than ⊢ but just plain testing on floating point values, since finite differencing is going to end up testing your primal function on floating point values anyway once you perturb it.
Though you might have a AD special case rule. (I don't think anyone really should)
JuliaDiff/FiniteDifferences.jl#188 (comment)