Sparse categorical accuracy with logits

This is a metric to use with Keras where:

• labels are indices corresponding to a class, and
• targets are logits

This is an iteration on Keras's built-in sparse_categorical_accuracy-metric to use with models where sparse_softmax_cross_entropy_with_logits-loss is used.