Add Tier I ECS diagnosis via Gregory regression

[duncanwp]

Summary

• Implements Tier I ECS diagnosis via Gregory regression on abrupt-4xCO2 / piControl experiments (closes Tier I: ECS diagnosis via Gregory regression (metadata) #77)
• Adds DataFinder.load_experiment_ds() for loading arbitrary CMIP6 experiments (piControl, abrupt-4xCO2, etc.) without the historical+SSP concatenation logic
• Hardens standardize_dims() to gracefully handle out-of-bounds time coordinates common in long control runs
• Adds rsdt (incoming TOA SW flux) to VARIABLE_FREQUENCY_GROUP and makes DataFinder handle variables without observation paths

New script: benchmark_scrips/ecs_benchmark.py

Loads tas, rsdt, rsut, rlut from piControl (baseline) and abrupt-4xCO2 (first 150 yr by default), performs linear regression of TOA net flux vs global-mean temperature, and outputs ECS / lambda / forcing / R-squared to results/ecs/ecs_results.csv.

Usage:
python ecs_benchmark.py --model CanESM5
python ecs_benchmark.py --model UKESM1-0-LL --n_years 150 --save_to_cloud

Test plan

• Run python ecs_benchmark.py --model CanESM5 and verify ECS output is in expected range (~5.6 K for CanESM5)
• Verify results/ecs/ecs_results.csv is created with correct columns
• Run with --save_to_cloud flag and verify GCS upload
• Confirm existing model_benchmark.py still works unchanged

Generated with Claude Code