Add ValorClassificationManager and refactor classification.py

[ntlind]

Improvements

• Refactor classification.py to do fewer joins and do them farther up the chain, allowing us to store less data in ValorClassificationManager. Results in a ~2 second speedup on all classification benchmarks.
• Implement ValorClassificationManager to allow users to add datums sequentially, rather than passing a full list
• Disallow passing empty lists of predictions to ValorClassificationManager and ValorDetectionManager. Add tests showing users can still pass a Prediction with an empty list of annotations.
• Split conftest into conftest_inputs and conftest_outputs. Move expected answers out of test files and into conftest_output.py
• Fixed a bug where the number of unique labels in a classification's metadata was being counted before the label map was applied, resulting in eval_job.meta["labels"] being too high for label map evaluations.