Update README for the new structure

[KyberNull]

This pull request updates the README.md to document new features and configuration changes, especially around large-raster inference and data structure. The updates clarify the workflow, add new configuration defaults, and update directory conventions for all stages (pretrain, train, evaluate, inference). It also documents the new inference entrypoint and output formats.

Major documentation and workflow updates:

• Added documentation for large-raster inference, including patch-wise TIFF inference and vectorized shapefile export, and described the new main.py entrypoint for this functionality. [1] [2] [3] [4]
• Updated and expanded configuration defaults for all stages (pretrain, train, eval, inference), including batch sizes, data roots, and new inference parameters.

Project structure and naming:

• Renamed phase_io.py to io.py in the training/ directory for clarity.

Data directory conventions:

• Standardized and clarified data directory layout and naming for pretrain, train, validation, and test datasets, including new evaluation and inference paths.

These changes ensure the documentation accurately reflects the current state of the codebase, improves clarity for new users, and introduces support for large-scale inference workflows.