Refactor DSAD for newer flux.jl explicit API

[rolling-robot]

I was tinkering around with Flux and anomaly detection and found that Flux changes its API starting from 0.13 to what they call "explicit style". So these are changes to support Flux 0.13 and above for DSAD algorithm.

More: https://fluxml.ai/Flux.jl/stable/training/training/#Model-Gradients and https://fluxml.ai/Zygote.jl/dev/#Explicit-and-Implicit-Parameters-1

My testing code:

using MLJ
using PyCall
using Flux
using OutlierDetection
using OutlierDetectionInterface: Labels, Data
using Plots
using CategoricalArrays

n_train = 200
n_test = 200

ocnn = @load DSADDetector pkg=OutlierDetectionNetworks verbosity=0

skl_ds = pyimport("sklearn.datasets")
data, labels = skl_ds.make_moons(n_train, noise=0.1)
fig = scatter(data[:,1], data[:,2], marker=:+)

anomaly_labels = map((x ->  "normal"),labels)
push!(anomaly_labels, "outlier")
data = vcat(data, [1.5;0.5] |> transpose)
push!(anomaly_labels, "outlier")
data = vcat(data, [-0.5;0.] |> permutedims)

encoder = Chain(
    Dense(2 => 4, relu, bias=false),
    Dense(4 => 8, relu, bias=false),
    Dense(8 => 15, relu, bias=false)
)
decoder = Chain(
    Dense(15 => 8, relu, bias=false),
    Dense(8 => 4, relu, bias=false),
    Dense(4 => 2, relu, bias=false))

loss_log = Vector()
detector = ocnn(
    encoder=encoder,
    decoder=decoder,
    epochs=120,
    callback = (
        ((m, x) -> ()),
        ((m, x, y) -> push!(loss_log, mean(m(x)))))
)

model, score = OutlierDetection.fit(detector,
    data |> permutedims,
    CategoricalArray(anomaly_labels, levels=["normal", "outlier"]),
    verbosity=2)

test = rand(Float64, (n_test,2)) * 11 - ones(n_test,2)*5
ŷ = OutlierDetection.transform(detector, model, test |> permutedims)
contour!(fig,
    range(-2, stop = 3, length = 200),
    range(-2, stop = 3, length = 200),
    (x, y) -> first(OutlierDetection.transform(detector, model, [x y] |> permutedims)),
         levels=30)

Noticeable changes:

1. DSADModel now contains only encoder and radius for hidden variables (decoder seemed to be redundant there)
2. DSADModel can be executed and returns svddScore for better integration with Flux.gradient
3. callback signature is changed. It is now tuple with two functions with different signatures.