predict errors for KNNRegressor when algorithm = :brutetree .

[OkonSamuel]

A user on julia discourse, recently reported that the code below fails during the prediction phase.

using DataFrames
using MLJ

X_rand = rand(50,3)
coef = [1; 0; 1]
y_rand = X_rand*coef + rand(50,1)

#convert to dataframe for MLJ inputs
X_rand_df = DataFrames.DataFrame(X_rand, :auto)
y_rand_df = DataFrames.DataFrame(y_rand, :auto)

train_idx, test_idx = partition(eachindex(y_rand_df[:,1]), 0.7, shuffle=true)

Tree = @load KNNRegressor
tree =Tree(algorithm=:brutetree)

mach = machine(tree,  X_rand_df[train_idx, :], y_rand_df[train_idx, 1])

MLJ.fit!(mach)

MLJ.predict(mach, X_rand_df[test_idx,:])

[OkonSamuel]

Further investigation shows that the above code also fails if KNNRegressor is replaced with any of the other exported KNN models, with algorithm=:brutetree.
This bug is because, although from the NearestNeighbors package, the :brutetree(BruteTree) constructor supports reorder and leafsize kwargs, they don't make sense in the predict phase. This is expected as BruteTree isn't actually a tree.