Fix test_approx() when one of the arguments is a broadcasted multidimensional array

[dfdx]

collect() on broadcasted arrays doesn't preserve its shape, but instead creates a flat array. This breaks tests e.g. for most activation functions in NNlib:

test_rrule(Broadcast.broadcasted, NNlib.σ, rand(3, 4))

which produces:

Test threw exception
  Expression: isapprox(actual, expected; kwargs...)
  DimensionMismatch("dimensions must match: a has dims (Base.OneTo(2), Base.OneTo(2)), b has dims (Base.OneTo(4),), mismatch at 1")

[codecov-commenter]

Codecov Report

Merging #230 (2c080a7) into main (63bbd48) will not change coverage.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##             main     #230   +/-   ##
=======================================
  Coverage   90.84%   90.84%           
=======================================
  Files          11       11           
  Lines         295      295           
=======================================
  Hits          268      268           
  Misses         27       27           

Impacted Files	Coverage Δ
src/check_result.jl	89.70% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 63bbd48...2c080a7. Read the comment docs.

[oxinabox]

Should this be fixed up stream in JuliaLang/Julia
JuliaLang/julia#43847 ?

[dfdx]

Amazing analysis! Let's see how it turns out in the Base and whether the fix gets to all popular Julia versions. If it's not applied to any of the versions supported in ChainRules, we can still apply this PR to be on the safe side.

[oxinabox]

Cool so this has now been addressed in julia v1.9.0
Unclear as to if this will be backported,
but we can add a comment to the source saying:
#Works around https://github.com/JuliaLang/julia/issues/43847 on pre-Julia v1.9
to this PR and then merge this if we still think this is needed?

[dfdx]

Yeah, since we support older versions of Julia, I believe it's still worth to have this workaround.
I added the suggested comment to the PR.