Delegate IColumn.fill_null/drop_null to Arrow (pytorch#96)

Summary:
Pull Request resolved: pytorch#96

As title

NOTE: the changes of diff only delegate IColumn.fill_null to Arrow since TorchArrow is using Apache Arrow 2.0.0, which doesn't support `drop_null` until Apache Arrow 6: https://arrow.apache.org/docs/python/api/compute.html#selections

Reviewed By: wenleix

Differential Revision: D32770009

fbshipit-source-id: 533df9e60582fe6f45266da6dac26576d6b0d5fb

torcharrow/icolumn.py

@@ -19,6 +19,7 @@
 
 from .dispatcher import Device
 from .expression import expression
+from .interop import from_arrow
 from .scope import Scope
 from .trace import trace, traceproperty
 
@@ -1134,12 +1135,10 @@ def fill_null(self, fill_value: ty.Union[dt.ScalarTypes, ty.Dict]):
         if not isinstance(fill_value, IColumn._scalar_types):
             raise TypeError(f"fill_null with {type(fill_value)} is not supported")
         if isinstance(fill_value, IColumn._scalar_types):
-            res = Scope._EmptyColumn(self.dtype.constructor(nullable=False))
-            for m, i in self._items():
-                if not m:
-                    res._append_value(i)
-                else:
-                    res._append_value(fill_value)
+            import pyarrow.compute as pc
+
+            arr = pc.fill_null(self.to_arrow(), fill_value)
+            res = from_arrow(arr, dtype=self.dtype, device=self.device)
             return res._finalize()
         else:
             raise TypeError(f"fill_null with {type(fill_value)} is not supported")