refactor: add rule for zero(::Any)

[sathvikbhagavan]

CI on master is broken. I suspect it is because of [email protected] - JuliaRegistries/General#99357

[sathvikbhagavan]

@ChrisRackauckas, the problem was coming from https://github.com/SciML/NeuralPDE.jl/blob/master/src/pinn_types.jl#L520 where it can be zero(::Vector{Float64}) which was erroring out like in https://github.com/SciML/NeuralPDE.jl/actions/runs/7625540660/job/20770044115#step:6:823.

I just added the rule here which was removed in JuliaDiff/ChainRules.jl#770 which fixes CI, although I am not completely sure if this is the right thing to do.

[ChrisRackauckas]

@MasonProtter @oxinabox I think that PR might not have been as non-breaking as you thought 😅.

[MasonProtter]

Ugh, this thing is a real whack-a-mole. I'm having trouble understanding though why zero(::Vector) should be non-differentiable and return NoTangent() instead of returning ZeroTangent(), that just seems wrong to me.

I'm curious @sathvikbhagavan, can you test what happens here if you try to do

using ChainRulesCore: @scalar_rule
@scalar_rule zero(x::Any) ZeroTangent()

instead of using @non_differentiable?

---

Edit: nevermind, I just remembered why that won't work, scalar rules are only for ::Number.

[MasonProtter]

Okay, this is now fixed by JuliaDiff/ChainRules.jl#773, so I think this PR is not needed.

[sathvikbhagavan]

Thanks @MasonProtter!