Non-diff shape handling in norm layers

[ToucheSir]

This reduces some latency when using Zygote.

using SnoopCompileCore
invalidations = @snoopr(using Zygote, Flux)
using SnoopCompile, AbstractTrees

loss(model, x) = sum(model(x))
tinf = let model = BatchNorm(3), x = rand(Float32, 16, 16, 3, 8)
    loss(model, x)
    @snoopi_deep gradient(loss, model, x)
end

julia> tinf
InferenceTimingNode: 2.846507/13.942681 on Core.Compiler.Timings.ROOT() with 139 direct children # master
InferenceTimingNode: 2.354729/9.970884 on Core.Compiler.Timings.ROOT() with 126 direct children # PR

PR Checklist

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