Add evaluate support for density estimators

[ablaom]

Merge conflicts will be best managed if this PR is merged after:

• Forthcoming update to bump the compat for CategoricalArrays to 1.0 (WIP of @ablaom)

---

Density estimators

This PR adds evaluate support for density estimators. According to the API docs such a
model is viewed as Probabilistic <: Supervised but provided X = nothing as
features. The target y are the samples to be fitted, and yhat = predict(model, nothing)
outputs a single pdf, the density being sought. A typical workflow using such a model might look like this:

y = <samples>
mach = machine(model, nothing, y) # `nothing` instead of features
fit!(mach)
yhat = predict(mach, nothing) # single pdf
log_loss(fill(yhat, length(y)), y)  # get an aggregate of the loss based on multiple samples, so need `fill` here

At present evaluate is unsupported because it tries to pair the single yhat with
multiple ground truth values y. The mitigation provided in this PR is to replace yhat with
FillArrays.Fill(yhat, nrows(y)) (a lazy version of fill(yhat, nrows(y))) in the case
that X == nothing. So then the following works:

evaluate(model, nothing, y, measure=log_loss)

After this PR is merged:

• Update [this link] so that it points to the example at the end of test/resampling.jl
  instead (some testing has been moved over in this PR).

cc @LucasMatSP

[LucasMatSP]

Small typo Estmiate