Keep ESANet working in 2025+ (#61)

- replace implementation for mIoU metric with pure PyTorch version to avoid dependencies on TensorFlow or PyTorch Ignite
- updated conda environment (Python 3.12 and PyTorch 2.9.1 -> env_esanet2025.yaml)
- add support for macOS with MPS backend (Python 3.12 and PyTorch 2.9.1 -> env_esanet2025_mac.yaml)
- fix lambda pickling issues when using multiple workers in DataLoader
- force that RandomCrop to always returns a contiguous array
- fix issues with RandomHSV augmentation (wrong range for value channel) -> might slightly affect training results
- minor dataset preparation instruction updates
- add arguments --toolbox_filepath and --data_filepath to SUNRGB-D preparation script to enable reusing already downloaded toolbox and data files
- switch to Apache 2.0 license
- add note to the corresponding PhD thesis

minor fix