Reimplement SAE training eval code

[Stevinson]

I have moved the evaluation code into an Evaluator class. The idea here was that this would be easier to extend and also that there are two different evaluation contexts, TrainingEval and PostTrainingEval. This was done so that a cheaper form of evaluation can be done during training, such that one can get an idea of training but one would want to perform a full evaluation on a model that is already trained.

• I've split evals.py into evaluator.py and eval_utils.py. I haven't changed the evaluation code itself.
• Specific evaluations can be ran by adding them to the EvalConfig