Feature/urban heat islands

[BachirNILU]

What does this PR do?

Adds the Guatemala Urban Heat Islands use case (WP330) and several
framework improvements required to support it.

Framework changes (general — affect all use cases):

New: Regression model and head

• Add src/models/predictive_model_regression.py — Lightning module
  for regression tasks (MSE loss, MAE/RMSE/R² metrics)
• Add src/models/components/pred_heads/mlp_regression_head.py — MLP
  prediction head for continuous targets

New: Multimodal encoder

• Add src/models/components/eo_encoders/multimodal_encoder.py —
  unified encoder replacing three separate encoders (coords-only,
  tabular-only, fusion). Controlled via use_coords and use_tabular flags.

Fix: BaseDataModule

• Add tabular_dim for "feat_" property
• Expose spatial_split_distance_m as configurable parameter
  (was hardcoded to 4000m)

Fix: BaseDataset

• Add support for feat_* tabular columns

Use case specific changes (Guatemala LST):

• Add scripts/make_model_ready_heat_guatemala.py — preprocessing
  script with NaN handling, convert categorical features and rename columns to "feat_*"
• Add src/data/heat_guatemala_dataset.py — dataset class
• Add experiment/model/data configs: fusion, coords-only, tabular-only regression

Before submitting

• Title is self-explanatory and description concisely explains the PR
• PR does only one thing — NOTE: bundles framework + use case changes,
  kept together as they are tightly coupled
• Breaking changes listed — none, all changes are additive
• Tested with pytest — pre-existing failure unrelated to this PR:
  ModuleNotFoundError: No module named 'pkg_resources'
  (setuptools missing in Python 3.12 env, exists on main branch too)

[BachirNILU]

CI tests were failing due to a pkg_resources issue in Python 3.12.
After investigating, I found this was already fixed in the develop branch
(see #46). Resolved by merging develop into this branch.

[gabrieletijunaityte]

Very good changes and nice documentation of changes!

I made small comments in the code, and I made a change to remove the save_dir argument from the parameters, as we have a consistent data directory structure.

Additionally, I wonder whether it would make sense to keep only one type of predictive model and decouple the metric calculation into a separate class that can be assigned via configuration? self.log accepts a dictionary of metrics, so it would be a clean implementation. @vdplasthijs, what do you think?

[vdplasthijs]

From @gabrieletijunaityte

Additionally, I wonder whether it would make sense to keep only one type of predictive model and decouple the metric calculation into a separate class that can be assigned via configuration? self.log accepts a dictionary of metrics, so it would be a clean implementation. @vdplasthijs, what do you think?

Yes good idea! Let's also discuss this in next meeting.