[ogb_lsc] the code for pcq does not use the train split during training

In the code for the dataset PCQM4M, the function [`_load_smiles`](https://github.com/deepmind/deepmind-research/blob/6fcb84268e74af981ae1496bfc2cb9ba9d701ef2/ogb_lsc/pcq/dataset_utils.py#L279-L298) is used to load data when building data iterator.

However, when the `split` is set to `train`, the function [`datasets.load_all_except_kth_fold_indices`](https://github.com/deepmind/deepmind-research/blob/6fcb84268e74af981ae1496bfc2cb9ba9d701ef2/ogb_lsc/pcq/dataset_utils.py#L290) only loads the validation indices except the k-th validation fold, and the train split is not used.

@alvarosg @saran-t