Your KNN is not a single model to make predictions but rather a RegressionPartitionedModel. Which means it has all the statistics of the cross validation and all individual models for each fold. If you want to make predictions you will have to select one of the trained models, as example:
y_pred_test(i, 1) = predict(KNN.Trained{1}, x_set);
I don't really understand why you would need this internal cross-validation since you're already looping to each data point. For me it would make more sense to simply train the model in all training data for each iteration:
KNN = fitcknn(x_train, y_train);
y_pred_test(i, 1) = predict(KNN, x_set);
Then you have only one model and can use the predict function.
