Inference tests do not catch problems with thunked inputs

[mzgubic]

Consider:

julia> my_id(x) = x
my_id (generic function with 1 method)

julia> function ChainRulesCore.rrule(::typeof(my_id), x)
           my_id_pb(ȳ) = (NoTangent(), ȳ)
           function my_id_pb(ȳ::AbstractThunk)
               precision = rand() > 0.5 ? Float64 : Float32
               return (NoTangent(), precision(unthunk(ȳ)))
           end
           return x, my_id_pb
       end

julia> test_rrule(my_id, 2.0)
Test Summary:                | Pass  Total
test_rrule: my_id on Float64 |    7      7
Test.DefaultTestSet("test_rrule: my_id on Float64", Any[], 7, false, false)

while really we should be catching that the thunked cotangents result in type unstable pullbacks. This may seem obscure but is quite relevant because of closures being unstable if they call another method of themselves.