Add WeightNorm reparametrization

[pxl-th]

Yet another attempt at adding WeightNorm.
Based on different bits found in #2053 [edited!] & #1005 with tests and documentation.

PR Checklist

• Tests are added
• Entry in NEWS.md
• Documentation, if applicable

[CarloLucibello]

This way we have duplication of the weights.
Can we let the original layer's weights act like v?
Something like

struct WeightNorm{which, dims, L, G}
    layer::L
    g::G
end

(w::WeightNorm)(x) = transform(w)(x)

function transform(wn::WeightNorm{which, dims}) where {which, dims}
    ϵ = eps(eltype(wn.v))
    v = getfield(wn.layer, which)     
    n2 = sum(abs2, v; dims)
    w = @. wn.g * v / sqrt(n2 + ϵ)
    fields, ctor = Functors.functor(wn.layer)
    return ctor(merge(
        fields, NamedTuple{(which,)}((w,)),
    ))
end

[codecov]

Codecov Report

Attention: Patch coverage is 76.92308% with 6 lines in your changes missing coverage. Please review.

Project coverage is 32.78%. Comparing base (2bbd8b3) to head (54319e1).
Report is 1 commits behind head on master.

Files with missing lines	Patch %	Lines
src/layers/normalise.jl	76.92%	6 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2550      +/-   ##
==========================================
+ Coverage   31.95%   32.78%   +0.82%     
==========================================
  Files          34       34              
  Lines        1987     1998      +11     
==========================================
+ Hits          635      655      +20     
+ Misses       1352     1343       -9     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[pxl-th]

Not yet ready. Needs some adjustments for GPUs

[pxl-th]

Added GPU tests as well, making a small test suite which we can exapnd in future to avoid duplication.
Should be ready for review

[CarloLucibello]

The doctest needs to be fixed. besides that looks good

[pxl-th]

All tests now pass.