fix: update tokenizer handling to avoid model config mismatch warning and align pad_token_id with eos_token_id

fix: use pad_token_id to avoid token registry warning, drop redundant dtype in from_pretrained

fix: use Image.add_local_file() — correct Modal 1.x API to embed dataset.py

fix: replace removed modal.Mount with Image.copy_local_file() for Modal 1.x

fix: mount modal/ dir so dataset.py is importable in container, expand cache cleanup to repo root

feat: restructure remote training files under modal/ and expand synthetic CBT dataset