Support for large number of classes when using tf.metrics.mean_per_class_accuracy() (#15946)

* Switched to using two 1-D variables instead of 2-D matrix for tf.metrics.mean_per_class_accuracy to reduce memory usage and allow larger number of classes.

* Fixed implementation and documentation of tf.metrics.mean_per_class_accuracy by:
- Casting labels to int64
- Handling weights correctly
- Changing return value of update op to per class accuracy tensor
- Updating outdated statements in documentation

Also removed reliance on confusion matrix in tests.