UNet COT-FM experiments: EMD/sinkhorn ablation + architecture search

[Justinezgh]

This PR some tasks of #14. It adds infrastructure and results for the UNet hyperparameter search:

• cot_fm.py: add --ot_method (emd/sinkhorn_log), --exp_config YAML loader, and post-training PQMass evaluation on last and best checkpoints
• scripts/ot_diagnostic.py: compute OT pairing quality (entropy, $E[|\Delta\theta|]$, displacement) without training to narrow eps/ot_reg candidates
• scripts/ps_unet_analysis.py: compute power spectrum and PDF of UNet-generated maps (via ODE solve) vs N-body reference
• configs: phase 1 (EMD, sinkhorn ot_reg=1e-3) and phase 2 (small/large arch) experiment YAMLs
• docs: experiment tracking doc with phase rationale, results, and figures

Phase 0+1 conclusion: EMD (χ²_best=113.6) outperforms sinkhorn; phase 2 architecture ablation (small/large) in progress.