Fix correctness issues found in full project review

Fix tokenizer loading: load from base model to avoid TokenizersBackend error

Pin transformers<5.0 and fix CPU model loading for HF Spaces

Add Phase 7: CPU fallback in infer.py and README with Spaces frontmatter

Add Phase 5: inference function (src/infer.py)

Timestamp evaluation result filenames to preserve multiple runs

Add Phase 4: ROUGE evaluation script and Colab notebook

Subsample train split to 4000 examples to fit Colab free T4 session

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

Fix LoRA targets for native Phi-3 and update notebook Hub link

Replace all SAMSum references with DialogSum across the project

Fix Phi-3 RoPE error, rename repo references to dialogue-summarizer

Fix hardcoded dataset name in CLI print statement

Switch dataset from Samsung/samsum to knkarthick/dialogsum

Fix SAMSum dataset path: samsum -> Samsung/samsum

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