[WIP] Don't rely on configs in prepare_for_gfn

[vict0rsch]

Not tested yet, feel free to do it yourself in the mean time.

This would be my preferred direction if we don't hit more configs-related obstacles.

[vict0rsch]

Make loading the data optional ; MAKE SURE this does not break GflowNet validation procedure which could expect the dataloaders

[vict0rsch]

Implemented a first version refactoring BaseTrainer.load(...) to use a new kwarg called prevent_load which is used in prepare_for_gfn as

config["prevent_load"] = {
    "logger": True,
    "loss": True,
    "datasets": True,
    "optimizer": True,
    "extras": True,
}

[lpodina]

On origin/catalyst_parallel for gflownet-dev and origin/faster_surface for ocp, and this command can trigger a permission denied error:
HYDRA_FULL_ERROR=1 python -W ignore main.py user=elena.podina +experiments=catalysts/asir_magpie logger.debug=True env.atomgraphconverter.n_cpu_threads=64 gflownet.optimizer.batch_size.forward=64 gflownet.optimizer.n_train_steps=3 +env.atomgraphconverter.sample_timeout=2 +env.atomgraphconverter.pyx_timeout=1