normalize model outputs for kNN evaluation

[clemsgrs]

In the kNN evaluation code, the class probabilities retrieved from model output don't sum to 1:

dinov2/dinov2/eval/knn.py

Lines 192 to 195 in e1277af

	def forward(self, features_dict, targets):
	for k in self.keys:
	features_dict = features_dict[k]
	return {"preds": features_dict, "target": targets}

This causes issues when computing kNN metrics.
I would suggest adding the following line of code to ensure probabilities sum to 1:

def forward(self, features_dict, targets):
    for k in self.keys:
        features_dict = features_dict[k]
    features_dict = features_dict / features_dict.sum(dim=-1).unsqueeze(-1)
    return {"preds": features_dict, "target": targets}