Add rule for Dict iteration

[ToucheSir]

Fixes the first example in #1065. N.B. that _pullback is used over @adjoint because we're trying to get rid of it, and over rrule because support for Dict tangents in CR is still spotty (+ we don't have to be nearly as general).

[CarloLucibello]

NNlib failure real but unrelated? And nightly as well.

[ToucheSir]

Nightly failure has been around for ages and should be fixed by #1280. NNlib failure I'm unsure of, need to investigate (Edit: not related, repros on NNlib master with tagged Zygote).

[ToucheSir]

CI looks green now that nightly and downstream failures have been addressed.

[mohamed82008]

What's blocking this PR?

[mohamed82008]

Any update here?

[oxinabox]

This looks right to me.
Except we should either mutate the gradient and return nothing
Or out of place compute the change in gradient and return it to be accumed later.
Not both.

But this is a problem with all the Dict code. Which is why accum(::Dict, ::Dict) is defined as a no-op

So approving

[ToucheSir]

Except we should either mutate the gradient and return nothing
Or out of place compute the change in gradient and return it to be accumed later.
Not both.

I actually tested this after #1288 and found it broke some tests, but I don't recall which ones. Shall we continue the discussion there?

[mohamed82008]

Thanks for the PR and the review everyone! Can we get a release?

[ToucheSir]

JuliaRegistries/General#67749