Fix compiled refits with empty table schemas (#1241)

# Summary

Fix compiled processor refits that produce an empty table schema.
Preserve the selected column names and semantic types while making
constructor metadata safe to emit in captured graphs.

# Description of the change

On PyTorch 2.14.0, Python 3.12.14, CPU/macOS, this public SDM workflow
fails on the third refit:

```python
import torch
from sdm import TableTensor
from sdm.processing import DropConstantColumns

processor = DropConstantColumns()
fit_transform = torch.compile(
    processor.fit_transform, backend="eager", fullgraph=False
)
for data in (
    torch.ones(3, 3),
    torch.arange(9, dtype=torch.float32).reshape(3, 3),
    torch.ones(3, 3),
):
    fit_transform(TableTensor.from_tensor(data))
```

The generated constructor call contains enum keys such as
`Stype.numerical` without binding `Stype`, raising `NameError: name
'Stype' is not defined`.

Pass semantic type keys as their string values across the opaque
table-constructor boundary, then restore the enum keys inside the
constructor. This preserves the public schema and fixes the permissive
compilation refit failure; strict fullgraph fitting with value-dependent
column selection remains separate.

# Validation

- Table and constant-column suites: 63 passed, 14 skipped on CPU PyTorch
2.14.0.
- External reproduction fails on unchanged main and passes with this
change.
- Repeated public refits with eager and Inductor backends preserve
empty, nonempty, and zero-width output schemas.
- Changed-file pre-commit checks passed, including formatting and type
checking.

Author: aw471

sdm/tensor/table.py

@@ -338,8 +338,9 @@ def __new__(
         if len(column_names) != len(set(column_names)):
             raise ValueError("Expected column names to be unique")
 
+        # Keep enum objects out of captured graph arguments.
         return TableTensor._make_table_tensor(
-            columns=columns,
+            columns={str(stype): names for stype, names in columns.items()},
             numerical=numerical,
             categorical=categorical,
             datetime=datetime,
@@ -352,7 +353,7 @@ def __new__(
     @staticmethod
     @torch.compiler.allow_in_graph
     def _make_table_tensor(
-        columns: dict[Stype, tuple[str, ...]],
+        columns: dict[str, tuple[str, ...]],
         numerical: Tensor,
         categorical: CategoricalTensor,
         datetime: Tensor,
@@ -361,9 +362,12 @@ def _make_table_tensor(
         cls: type[TableTensor] | None = None,
     ) -> TableTensor:
         cls = TableTensor if cls is None else cls
+        stype_columns = {
+            Stype(stype): names for stype, names in columns.items()
+        }
         column_to_loc = {
             name: (stype, i)
-            for stype, names in columns.items()
+            for stype, names in stype_columns.items()
             for i, name in enumerate(names)
         }
         out = Tensor._make_wrapper_subclass(
@@ -378,7 +382,7 @@ def _make_table_tensor(
         out._datetime = datetime
         out._text = text
         out._id = id
-        out._columns = columns
+        out._columns = stype_columns
         out._column_to_loc = column_to_loc
         return out