Feat/data pipeline

[Thomson-Lam]

Changes

• Updated data pipeline to ingest and clean the Albert Historical Wildfires Dataset instead of live data from CWFIS.
• Removed FIRMS and CWFIS ingestion and kept CFFDRS ingestion.
• Removed XGBoost model which was originally trained on synthetic data for fire spread rate and replaced with observed spread rate from the dataset.
• Added a lightweight cleaning process in the data pipeline for simple vetting and reporting missing or invalid essential values for the environment params.

Validation:

Ran uv run python -m src.ingestion.static_dataset --target-count 50000 --cffdrs-year 2025 --raw-alberta-csv data/static/fp-historical-wildfire-data-2006-2025.csv and ingested ~2000 samples with some enriched with corresponding geographically approximated CFFDRS data.

Noted that missing CFFDRS data can introduce bias for the RL agent training and would make the paper harder to defend, so elected to drop the --cffdrs-year flag for ingesting the final frozen data for training in the README instructions for usage, and to just use the Alberta dataset instead, which already contains enough core signals. Updated docs/data-pipeline.md and README.md for usage and reflected changes.