Migrate to fastai v2 and modernise project tooling

[mnpinto]

Breaking changes

• FRP preprocessing: log1p is now applied to FRP values at dataset creation time (via ActiveFiresLog) rather than at load time. Datasets created with the old pipeline (ActiveFires) are not compatible — they must be regenerated with the new banet_create_dataset command.
• Data format: predictions and masks are now stored as .nc (NetCDF4) instead of .mat. Existing .mat datasets cannot be used directly with this version.

Major changes

Core

• Migrated all modules (data.py, predict.py, nrt.py, train.py, cli.py, etc.) from fastai v1 to fastai v2
• Data format updated from .mat to .nc (NetCDF4); active fires now preprocessed with log1p via ActiveFiresLog before saving
• VIIRS 750m support with updated download and preprocessing pipeline

Training (new)

• Added full banet/train.py with SampleEpisode, CustomLoader, ImageSequence, CutoutCombined, BCE, and train_model()
• Added banet_train_model CLI command

Tooling

• Replaced setup.py + settings.ini with pyproject.toml
• Removed nbdev dependency; source notebooks deleted (.py files are now the source of truth)
• Simplified Makefile: make sets up the venv, make test runs pytest
• CI updated to Python 3.11 and actions/setup-python@v5
• Replaced fastscript with fastcore.script

Cleanup

• Removed all AUTOGENERATED headers and unused nbdev imports from source files
• Deleted banet/_nbdev.py
• Updated CONTRIBUTING.md and README.md

Contributors

• @timbelime — contributed the updated train.py implementation (fastai v2 training pipeline) via PR Timbelime train v2 #14