Implement sample-wise accuracy to check success of multiple attacks

[maurapintor]

The new class would add a new metric that takes a list of dataloader (resulting from multiple attack runs) and produce a compound metric that evaluates the worst-case prediction over the trials.

The trick is to collect all y_preds from the different loaders, and when computing the accuracy, add a minimum over the "correctness" of the predictions.

Example:

# stack predictions vertically
y_pred = [
    [1, 2, 3],
    [1, 2, 5],
    ]
y_true = [1, 2, 3]

# expected result:  [correct, correct, wrong]

This is the implementation:

correct = (y_pred.type(y_true.dtype).cpu() == y_true.cpu())

# result: 
# [[1, 1, 1],
#  [1, 1, 0]]

# take worst case
correct = correct.min(dim=0).values

# result: [1, 1, 0] -> sum for accuracy

A new metric will be added in metrics.classification.