refactor: modernize codebase for current Python & PyTorch stack

[Omdeepb69]

Summary

This PR modernizes the legacy NMT pipeline to make it compatible with the current Python and PyTorch ecosystem while preserving the original architecture and functionality.

Key Changes

• Updated deprecated PyTorch APIs to the current stable equivalents
• Refactored training and evaluation modules for compatibility with modern tensor operations
• Cleaned up legacy code patterns and improved overall readability
• Updated requirements.txt with working dependency versions
• Added .gitignore to remove cache artifacts from version control
• Ensured preprocessing, training, and inference scripts run on a modern environment

Why this is needed

The existing implementation depends on older Python/PyTorch versions and cannot run in a current environment without significant fixes.
This PR enables:

• Reproducibility on modern systems
• Easier onboarding for new contributors
• Future research and experimentation on top of this codebase

Backward compatibility

The core model logic and workflow remain unchanged.
Only compatibility and maintainability improvements were introduced.

Testing

• Verified preprocessing pipeline
• Verified training script execution
• Verified translation/inference flow

Future Work (optional)

• Add Docker / reproducible environment
• CI for training smoke test
• HuggingFace dataset/tokenizer integration