One point datasets always end up in validation split

Created by: Zaharid

When the dataset has one data point and a frac < 1 the code here

https://github.com/NNPDF/nnpdf/blob/43e7e752f04852bd9f2b458e6d03fb088586e762/validphys2/src/validphys/n3fit_data.py#L95

rounds the number of training points to zero, and so the dataset end in validation for all replicas. This may cause some total cross sections like ATLASTTBARTOT7TEV to end up with a somewhat worse chi2.