Wrong function is used for calculating dense score

[dangkhoasdc]

In the doc of dense_scores, it is mentioned that: torch.mean is dedicated to training

neural-cherche/neural_cherche/utils/dense_scores.py

Lines 138 to 140 in 2df0214

	func
	Either torch.sum or torch.mean. torch.mean is dedicated to training and
	torch.sum is dedicated to inference.

However, in your training function of SparseEmb, you use torch.sum instead. So which one is preferable for training? I guess it should be torch.mean?

neural-cherche/neural_cherche/train/train_sparse_embed.py

Line 129 in 2df0214

func=torch.sum,