Update configs

[sgreenbury]

This pull request adds and updates several configuration files and documentation to support a new, centralized batch of planned ablation studies on the CNS (Conditioned Navier-Stokes) task. The main focus is on adding ablation configs for model architecture, loss variants, noise channels, and diffusion, as well as updating documentation to reflect the new planned batch and its implementation status. A new evaluation script for ensemble size ablation is also included.

The most important changes are:

New experiment configurations:

• Added CNS ablation configs for U-Net architecture (crps_unet_azula_80m.yaml), CRPS loss variants (FairCRPS and plain CRPS), noise channels (crps_vit_noise256.yaml), and diffusion ViT baseline (diffusion_vit_large.yaml). These configs standardize architecture, loss, and training parameters for fair comparison. [1] [2] [3] [4] [5]

Documentation updates for planned CNS batch:

• Updated slurm_scripts/ablations/README.md to describe the new planned CNS batch, list its runs, and mark the implementation status of each ablation. Added a section explaining the batch submission scripts and how each config/run fits into the overall study. [1] [2]
• Updated architecture-specific READMEs to reflect the new U-Net config, clarify the status of FNO, and document parameter-matching decisions. [1] [2]

Loss variant and ensemble size documentation:

• Marked the CRPS loss variant configs as implemented and clarified class paths for loss/metric targets. Updated the ensemble size ablation docs to include the planned n_members=4 ViT run with batch size details. [1] [2] [3] [4]

Ablation status and evaluation scripts:

• Updated the status and evaluation details for cached-latent CRPS and global conditioning ablations, including instructions for reruns and eval script locations. [1] [2] [3]
• Added a new eval script to select the best multi-Winkler checkpoint after 0.25 progress for ensemble size ablation runs.