train_valid_split sometime fails.

[corochann]

BaseSplitter's train_valid_split checks assert len(test_inds) == 0.

when i use RandomSplitter, it failed. i guess calculation of

        train_data_size = int(len(dataset) * frac_train)
        valid_data_size = int(len(dataset) * frac_valid)

may sometimes remain rest indices even if frac_train + frac_valid == 1.

we should check other splitter class and add test too.