optimization with flex-attention and packed sequence and graph2sequence generation

1. Major performance optimizations:
   • Flex Attention: Implemented flex_attention with sequence packing for efficient multi-sample training, achieving significant speedup
   • Speed improvements: AI-optimized Eulerian path algorithms, reduced CUDA sync overhead, tokenization pipeline optimizations
   • SDPA support: Added scaled dot-product attention with automatic validation fallback
2. Monitoring & profiling: Integrated WandB experiment tracking and Torch Profiler for performance analysis
3. Code quality: Major tokenizer refactoring (modular design), removed redundant parameters, added pre-commit enforcement