Fix loading of pooling layers

[adrhill]

As discussed on Slack, I tried my hand at fixing #2584 and FluxML/Metalhead.jl#287.
This line of code appears to be enough to make Metalhead's VGG models load.

Without it, the following error is thrown:

julia> using Metalhead

julia> VGG(19; pretrain=true)
ERROR: ArgumentError: Tried to load Base.OneTo(0) into (:pad, :k, :stride) but the structures do not match.
Stacktrace:
  [1] loadmodel!(dst::MaxPool{2, 4}, src::Tuple{}; filter::Function, cache::IdSet{Any})
    @ Flux ~/Developer/Flux.jl/src/loading.jl:104
  [2] loadmodel!(dst::Tuple{…}, src::Tuple{…}; filter::Function, cache::IdSet{…})
    @ Flux ~/Developer/Flux.jl/src/loading.jl:118
  [3] loadmodel!(dst::Chain{Tuple{…}}, src::@NamedTuple{layers::Tuple{…}}; filter::Function, cache::IdSet{Any})
    @ Flux ~/Developer/Flux.jl/src/loading.jl:118

I could use some guidance on whether this fix is adequate and how to test it.

[mcabbott]

To be clear the before-and-after states are

julia> Flux.state(MaxPool((2,3)))
()

(jl_gaQBh1) pkg> st Flux
Status `/private/var/folders/yq/4p2zwd614y59gszh7y9ypyhh0000gn/T/jl_gaQBh1/Project.toml`
⌃ [587475ba] Flux v0.14.25

and

julia> Flux.state(MaxPool((2,3)))
(k = (2, 3), pad = (0, 0, 0, 0), stride = (2, 3))

(@v1.11) pkg> st Flux
Status `~/.julia/environments/v1.11/Project.toml`
  [587475ba] Flux v0.16.3

So this looks fine?

• Could use a comment saying what it's for!
• Should probably be in https://github.com/FluxML/Flux.jl/blob/master/src/deprecations.jl
• Might be worth getting a few more layers which will have the same problem, e.g. https://fluxml.ai/Flux.jl/dev/reference/models/layers/#Pooling

[adrhill]

For the sake of documentation:

Flux v0.14:

julia> Flux.state(MaxPool((5, 5)))
()

julia> Flux.state(MeanPool((5, 5)))
()

julia> Flux.state(AdaptiveMaxPool((5, 5)))
()

julia> Flux.state(AdaptiveMeanPool((5, 5)))
()

julia> Flux.state(GlobalMaxPool())
()

julia> Flux.state(GlobalMeanPool())
()

Flux v0.16:

julia> Flux.state(MaxPool((5, 5)))
(k = (5, 5), pad = (0, 0, 0, 0), stride = (5, 5))

julia> Flux.state(MeanPool((5, 5)))
(k = (5, 5), pad = (0, 0, 0, 0), stride = (5, 5))

julia> Flux.state(AdaptiveMaxPool((5, 5)))
(out = (5, 5),)

julia> Flux.state(AdaptiveMeanPool((5, 5)))
(out = (5, 5),)

julia> Flux.state(GlobalMaxPool())
()

julia> Flux.state(GlobalMeanPool())
()

So I guess the global pooling layers require no modification?