Prediction at endpoint yields NaN

[Skumin]

The prediction using Loess.predict fails at endpoints for (some) cases where the covariate is equally spaced:

using Distributions
import Loess

x = collect(1:50);
y = x .+ rand(Normal(0, 10), 50);

x = convert(Array{Float64}, x);
model = Loess.loess(x, y);
prds = Loess.predict(model, x)
# prds[1] is NaN

[tlnagy]

Duplicate of #28