Add ViT MAE pretrain ablation

[sgreenbury]

Summary

• add CNS deterministic ViT MAE/no-ensemble pretrain preset
• add timing and 24h MAE Slurm submitters with all progress checkpoints retained and W&B checkpoint artifact logging
• add MAE-to-CRPS timing and short fine-tune submitters using n_members=16 and batch_size=16
• document workflow for using MAE checkpoints to seed later CRPS fine-tuning

Validation

• bash -n slurm_scripts/ablations/vit_mae_pretrain/submit_vit_mae_pretrain_timing.sh
• bash -n slurm_scripts/ablations/vit_mae_pretrain/submit_vit_mae_pretrain_large.sh
• bash -n slurm_scripts/ablations/vit_mae_pretrain/submit_vit_mae_to_crps_timing.sh
• bash -n slurm_scripts/ablations/vit_mae_pretrain/submit_vit_mae_to_crps_large.sh
• uv run autocast epd --mode local --dry-run datamodule=conditioned_navier_stokes local_experiment=ablations/vit_mae_pretrain/conditioned_navier_stokes/vit_azula_large_mae_no_ensemble logging.wandb.enabled=false trainer.max_epochs=1
• uv run python -m autocast.scripts.train.encoder_processor_decoder --cfg job datamodule=conditioned_navier_stokes local_experiment=ablations/vit_mae_pretrain/conditioned_navier_stokes/vit_azula_large_mae_no_ensemble logging.wandb.enabled=true logging.wandb.log_model=all +trainer.max_epochs=1
• uv run autocast epd --mode local --dry-run datamodule=conditioned_navier_stokes local_experiment=epd/conditioned_navier_stokes/crps_vit_azula_large model.n_members=16 datamodule.batch_size=16 +resume_from_checkpoint=/tmp/fake_mae.ckpt +resume_weights_only=true logging.wandb.enabled=true logging.wandb.log_model=all +trainer.max_epochs=1
• uv run python -m autocast.scripts.train.encoder_processor_decoder --cfg job datamodule=conditioned_navier_stokes local_experiment=epd/conditioned_navier_stokes/crps_vit_azula_large model.n_members=16 datamodule.batch_size=16 +resume_from_checkpoint=/tmp/fake_mae.ckpt +resume_weights_only=true logging.wandb.enabled=true logging.wandb.log_model=all +trainer.max_epochs=1