Add CrackSAM support and enable multi-GPU training

[keyprocedure]

Summary

This PR integrates CrackSAM support, enables distributed training, and adds associated framework improvements. These changes are bundled into a single PR because of their interdependence and the pace of this sprint.

CrackSAM:

• Integrates SAM (Segment Anything Model) encoder.
• Adds custom CrackSAM loss.
• Adds support for multi-channel model outputs.

Framework Enhancements:

• Adds PyTorch DDP support enabling multi-GPU training.
• Adds per-model img_size config to support dynamic dataset resolutions.
• Corrects loss selection for non-Dice loss configurations.
• Adds support for the Khanhha dataset.
• Updates usage documentation.
• Removes epoch loss printing during training.

Fixes #34
Fixes #43
Contributes to #3, #14, #42

Test Plan

Verified complete training, validation, and test pipeline using:
python -m main -c cracksam.yml --dataset khanhha

Verified multi-GPU training on 4×H100 GPUs with:
torchrun --nproc_per_node=4 main.py -c cracksam.yml --dataset khanhh