Add stations/era5_cell_splits.json, the split by ERA5 grid cell
Adds stations/era5_cell_splits.json: a second station split, in the same format as stations/quality_splits.json (PR #2), in which whole ERA5 grid cells are assigned to one group. Every station is assigned to its nearest ERA5 0.25-degree node, and all stations of a node share the group, so a test station never has a train, val or backbone neighbour in its own cell. This is the split used to train and evaluate the published Marmot model (its split_by_era5_node config), and it is the one to use when the evaluation is "held-out ERA5 cells".
Sizes: backbone 8,180, train 2,260, val 854, test 555 (the same sizes as the quality split, but a different assignment; the two agree only at chance level). The four lists are disjoint and cover all 11,849 station ids of stations/locations.geojson.
The dataset card (PR #1) does not mention this file yet; it will be updated once we decide how to present the two splits.