Add subseasonal calibration pipeline

[nefrathenrici]

Purpose

This PR adds a subseasonal calibration pipeline using ERA5 observations. The pipeline runs September months from 2018 to 2021 with the wxquest_diagedmf.yml config, collecting monthly surface flux diagnostics. The observations consist of ERA5 surface fluxes (latent heat, sensible heat, evaporation) with a latitude-weighted, diagonal scalar covariance structure. The only parameter currently calibrated is entr_inv_tau (inverse entrainment timescale). The implementation uses a modular structure with separate files for observation generation, loss functions, model interface, and calibration orchestration.

To-do

• Add tests
• Generate observations before running the pipeline

Content

• Introduces CalibrateConfig struct (from @ph-kev) to manage calibration settings (minibatch size, iterations, date ranges, spinup/extend periods)
• Moves calibration code from root to experiments/calibration/coarse_amip/ directory
• New files: generate_observations.jl, model_interface.jl, observation_map.jl, observation_utils.jl
• Removes the existing perfect model pipeline.

---

• I have read and checked the items on the review checklist.