Support reading sharded files

PGLearn is hosted on HF which has 50GB file limits. For the medium/large cases, some files exceed this limit, so splits are pushed instead, i.e. `SOCOPF/dual.h5` is replaced by a folder `SOCOPF/dual` containing files `xaa` `xab` which when `cat`'d together give the original `SOCOPF/dual.h5`. ML4OPF should merge these automatically upon first read (like how the HF dataset loading script does [here](https://huggingface.co/datasets/PGLearn/PGLearn-Large-9241_pegase/blob/script/PGLearn-Large-9241_pegase.py#L420-L426))