Add safetensors to the pytorch extra

[manjunathshiva]

A fresh pip install tabfm[pytorch] cannot load the released checkpoint. On tabfm 1.0.1 (cb6ba46), Python 3.13, torch 2.13.0:

from tabfm.src.pytorch import tabfm_v1_0_0
model = tabfm_v1_0_0.load(model_type="classification")

File ".../huggingface_hub/hub_mixin.py", line 813, in _load_as_safetensor
    if packaging.version.parse(safetensors.__version__) < packaging.version.parse("0.4.3"):
NameError: name 'safetensors' is not defined

Cause: huggingface_hub declares safetensors as an optional dependency and imports it lazily, but the PyTorch loader always goes through PyTorchModelHubMixin, and since 1.0.1 the Hugging Face checkpoint ships only model.safetensors (#33 / the 1.0.1 weight-loading fix dropped the pytorch_model.bin path). So safetensors is a hard runtime dependency of the pytorch extra — and torch does not pull it in transitively. Anyone installing per the README's PyTorch instructions hits the NameError on their first load().

It likely went unnoticed because CI's pip install -e .[dev,jax,pytorch] never loads the released checkpoint (tests build small models directly), and any environment that has ever installed transformers/accelerate already has safetensors present.

Change: add safetensors to the pytorch extra in pyproject.toml, and a matching pinned entry (safetensors==0.8.0, current release) to the requirements.txt lock in pip-compile format/ordering. The JAX path is unaffected — its weights restore via orbax.

Verified in a clean venv: pip install -e .[pytorch] on this branch pulls safetensors and tabfm_v1_0_0.load(model_type="classification") succeeds; on main the same steps reproduce the NameError above.

Note on overlap with #49: the MLX backend PR bundles this same dependency fix (both the safetensors line in the pytorch extra and the safetensors==0.8.0 pin in requirements.txt). This PR extracts it as a standalone fix so it isn't gated on the larger feature review; if this lands first, the corresponding hunks in #49 resolve trivially on its rebase, and if #49 lands first this PR can simply be closed.