Inference path imports datasets via re-export from dataset.dataset

`import needle` pulls in HuggingFace `datasets` (and its `pyarrow`/`pandas`/`fsspec` tree) even though nothing on the inference path uses it. The tokenizer symbols needed by `run.py` and `__init__.py` are defined in `needle/dataset/tokenizer.py`, but are imported through `needle/dataset/dataset.py`, which imports `datasets` at module level.

### Reproduction

Commit `ffb1c51`:

```bash
python -c "import sys; import needle; print('datasets' in sys.modules)"
# True
```

### Detail

Three sites reach `datasets` on the inference path:

- `needle/__init__.py:14`: `from needle.dataset.dataset import get_tokenizer`
- `needle/model/run.py:11`: `from ..dataset.dataset import get_tokenizer, to_snake_case, DEFAULT_MAX_ENC_LEN, DEFAULT_MAX_GEN_LEN`
- `needle/dataset/__init__.py:2`: `from .dataset import *`

`needle/dataset/dataset.py:15` then imports `datasets` unconditionally at module level.

All four symbols are defined in `needle/dataset/tokenizer.py` and merely re-exported by `dataset.py:19`. `tokenizer.py` itself only needs `sentencepiece` and `tqdm` at module level; `architecture.py` and `constrained.py` have no relative imports at all. So these three lines are the only coupling between inference and the training stack.

The third site matters most: `needle/dataset/__init__.py` executes on *any* `needle.dataset.*` import, so fixing only the first two has no effect.

### Why it matters

`generate()` / `load_checkpoint()` are usable as an inference dependency downstream, but importing them currently requires the training stack to be installed and imported at startup. For a 26M model aimed at constrained environments, that's a real cost for code the inference path never touches.

I appreciate this may be intentional if the Python package is considered a training/research harness with Cactus as the real inference runtime, happy to close if so. But since the fix is behavior-preserving and changes no dependencies or public API, it seemed worth raising.

Start with the import sites in needle/__init__.py, needle/model/run.py, and needle/dataset/__init__.py, then inspect needle/dataset/dataset.py and needle/dataset/tokenizer.py to trace the re-exports. Run the provided Python reproduction before and after the change; done means importing needle no longer places datasets in sys.modules while the listed tokenizer symbols remain available.

Written by the indexing model from the issue text.