Fix correctness issues found in full project review

Reduce batch size to 2, increase gradient accumulation to 8 for faster steps

Replace all SAMSum references with DialogSum across the project

Phase 1-3: data pipeline, LoRA model setup, and training loop