Skip data version checks unless evaluating

[jordis-ai2]

Running python scripts/datagen/run_pipeline.py --seed 3 with some data versions (e.g., scenes) different from the ones used for release triggers a pydantic error, which should only happen when evaluating.

[jordis-ai2]

This seems no longer to be an issue?

Setting, e.g.

        "ithor": "20251116",  # "20251217_with_occupancy",

in molmo_spaces/molmo_spaces_constants.py and calling python scripts/datagen/run_pipeline.py --seed 3 --scene_dataset ithor doesn't seem to trigger the error.

Idem with

        "procthor-objaverse-train": "20251217",  # "20251205_with_occupancy",
        "procthor-objaverse-val": "20251217", #  "20251205_with_occupancy",

and calling python scripts/datagen/run_pipeline.py --seed 3 --scene_dataset procthor-objaverse , or with

        "holodeck-objaverse-train": "20251205",  # "20251217_with_occupancy",
        "holodeck-objaverse-val": "20251205",  # "20251217_with_occupancy",

and calling python scripts/datagen/run_pipeline.py --seed 3 --scene_dataset holodeck-objaverse , or with

        "procthor-10k-train": "20251102", # "20251122_with_occupancy",
        "procthor-10k-val": "20251102", # "20251217_with_occupancy",
        "procthor-10k-test": "20251102", # "20251121_with_occupancy",

and calling python scripts/datagen/run_pipeline.py --seed 3 --scene_dataset procthor-10k.

[jordis-ai2]

In addition to the now obsolete issue initially described above, the resource manager initialization happens way ahead of time in some scripts. Working on this now.