Fix computing of array element type for ∇eachslice

[BioTurboNick]

Fixes #807

Added test as well. Two lines really belong in ChainRulesCore, not sure what the process for that is.

[BioTurboNick]

Other than the formatter, all the failures seem to be due to the fact that the changes mean Julia currently isn't fully inferring the output anymore. I'll see if there's anything to do about that.

[BioTurboNick]

Switching to Base.promote_eltype addresses the inference issue, but that's not a public API in base.

[BioTurboNick]

Huh. I'm not sure why the inference test is failing on 1.6 only. It's inferring Union{NoTangent, Vector{Float64}) instead of just Vector{Float64}.

[simeonschaub]

I believe the original intent with the design of similar was to support cases like Vector{Vector{Float64}} where zero(Vector{Float64}). That of course doesn't work though since similar will just produce #undefs here making this fail anyways.

Looks good to me minus the methods that need to be moved to ChainRulesCore. Could you make a PR there (ideally with some unit tests), so we can tag a new version there and then change the compat here.

[simeonschaub]

Just needs to drop the changes already in JuliaDiff/ChainRulesCore.jl#682 and bump the compat for ChainRulesCore

[BioTurboNick]

done and done

[simeonschaub]

Hmm, looks like the format suggestions are not working properly (Maybe because it's on a fork?) If you look at the CI logs, you can see that it complains about the formatting of line 257 in the tests. Once that's fixed this looks good to go!

[devmotion]

(Maybe because it's on a fork?)

Yes, due to restricted permissions for forks the suggestions can only be shown in the logs but not as actual suggestions in the PR.

[simeonschaub]

Thanks for the awesome contribution!