Ignore Zeros when gathering parameters

[pxl-th]

This fixes #1277.
IMO, it makes more sense to ignore disabled parameters when gathering model parameters, than have something that is not used and may lead to confusion.

PR Checklist

• Tests are added
• Entry in NEWS.md
• Documentation, if applicable
• Final review from @dhairyagandhi96 (for API changes).

[DhairyaLGandhi]

Removing Zeros from the list params might give incorrect results for loading complete chains where it is expected to see the bias term appear.

Another way of looking at it is, if you see an array where you try loading it with nothing (instead of zeros) you break the assumption that it might be populated with legitimate values if you train it further. If you see Flux.Zeros in the bias, ignore it, if you see an array either convert it to Flux.Zeros or load it with zeros.

[pxl-th]

In this case there is an issue, if we have a model with all biases disabled:

m = Chain(
    Conv((3, 3), 1=>2, bias=Flux.Zeros()),
    Conv((3, 3), 2=>3, bias=Flux.Zeros()),
    Conv((3, 3), 3=>4, bias=Flux.Zeros()),
)

Then in parameters we have Flux.Zeros only from first convolution, since parameters are stored in Zygote.Params which wraps IdSet{Any} and so we effectively loose information about all other disabled biases.

for p in Flux.params(m)
    println(typeof(p))
end

>>> Output:
Array{Float32,4}
Flux.Zeros{Bool,0}
Array{Float32,4}
Array{Float32,4}

[DhairyaLGandhi]

The crucial thing to note here is that we check for the presence of Flux.Zeros in the model side. That way the information is preserved. To make them different, we could make the struct mutable to create new instances of it with every call.

[pxl-th]

Closing as it is not relevant anymore