add PyTorchModelHubMixin to TabFM

[kashif]

Makes TabFM extend PyTorchModelHubMixin so it gets from_pretrained,
save_pretrained, and push_to_hub for free.

• load() now calls TabFM.from_pretrained(HF_REPO_ID, subfolder=model_type)
  instead of snapshot_download + manual torch.load + load_state_dict
• save_pretrained writes model.safetensors (preferred over .bin) plus a
  proper config.json with all init params including is_classifier
• _from_pretrained translates the legacy hub task: "classification" field to
  is_classifier: bool so existing weights load without any config update
• Removes the Config/ClassificationConfig/RegressionConfig dataclasses and
  all the manual json/bin saving from convert_and_upload.py

Users can now also do:

from tabfm.src.pytorch.model import TabFM
model = TabFM.from_pretrained("google/tabfm-1.0.0-pytorch", subfolder="classification")
model.push_to_hub("my-org/my-tabfm-fork")

[erzel]

Thanks for the contribution. Please see my comments below.

[kashif]

@erzel addressed all the review points: TabFM_HF now lives in tabfm_v1_0_0.py, subclasses TabFM and PyTorchModelHubMixin, and always delegates to the superclass _from_pretrained (only the config fetching differs by resolving to a local dir first). Duplicated config translation logic is merged into one helper, the pytype annotation is back, added a logging.warning when config.json is missing, and the local path check in load() raises immediately. Verified hub subfolder load, local dir load, and load() all work.

[erzel]

Thanks for making the changes.