Output embeddings to a geopandas.GeoDataFrame with columns 'source_url', 'date', 'embeddings', and 'geometry'. Essentially copying and adapting the code from a767164 in #73, but modifying how the encoder's masking is disabled, and how the mean/average of the embeddings is computed over a slice of the raw embeddings.

The output GeoParquet file now has a filename with a format like "{MGRS:5}_{MINDATE:8}_{MAXDATE:8}_v{VERSION:3}.gpq", e.g. "12ABC_20210101_20231231_v001.gpq". Have implemented this in model_vit.py, and copied over the same `on_predict_epoch_end` method to model_clay.py. Also, we are no longer saving out the index column to the GeoParquet file.

Forgot to update the filename in the unit test to conform to the new `{MGRS}_{MINDATE}_{MAXDATE}_v{VERSION}.gpq` format. Patches f19cf8f.

Splitting the previous integration test on the neural network model into separate fit and predict unit tests. Only testing the prediction loop of CLAYModule, because training/validating the model might be too much for CPU-based Continuous Integration. Also for testing CLAYModule, we are using 32-true precision instead of bf16-mixed, because `torch.cat` doesn't work with float16 tensors on the CPU, see pytorch/pytorch#100932 (should be fixed with Pytorch 2.2).

Decided that the index column might be good to keep for now, since it might help to speed up row counts? But we are resetting the index first before saving it. Partially reverts f19cf8f.

After f1439e3, need to ensure that the index column is checked in the output geodataframe.