[microTVM] Rework evaluate_model_accuracy into a more generic helper function

[guberti]

Currently, evaluate_model_accuracy addresses only a very specific use case - when one only cares about the average accuracy and runtime of a host-driven AOT model. This prevents users from implementing workarounds to bugs like #12538, and means we redundantly return both model predictions and the accuracy.

This PR changes evaluate_model_accuracy in python/tvm/micro/testing/evaluation.py into predict_labels_aot. This function has the type signature:

def predict_labels_aot(session, aot_executor, input_data, runs_per_sample=1):

Unlike evaluate_model_accuracy, predict_labels_aot does not take as input the true labels of the data. Instead, it returns an iterator of (prediction, runtime) tuples, and lets users do what they want with these.

cc @alanmacd @gromero @mehrdadh