Model testing example for WebNNM framework

Neural network models are evaluated using a test set of features and their classifications. A given dataset is first split into training and evaluation subsets. The evaluation graph extends the WebNN computational graph for inference with a few further operators to compute the average loss per sample. For large datasets, the evaluation samples are further split into batches for greater computational efficiency when compared to evaluating single samples. The last batch will usually be a little smaller than the rest. WebNNM implements a cache of WebNN graphs that are built on demand for the actual batch size.

This example demonstrates the use of a simple classification model on a tiny dataset for 3 species of iris flowers. The dataset is represented as a CSV file with 4 numeric features for each species.