Fix .where for sparse blocks

Discrepancy comes from:

	 dense_frame._data.blocks[0].values  # this is 2D even for 1D block
	sparse_frame._data.blocks[0].values  # this is always 1D

I'm sure this had worked before and was unneeded a month ago.

Author: kernc

pandas/core/internals.py

@@ -1456,6 +1456,11 @@ def where(self, other, cond, align=True, errors='raise',
             raise ValueError("where must have a condition that is ndarray "
                              "like")
 
+        # For SparseBlock, self.values is always 1D. If cond was a frame,
+        # it's 2D values would incorrectly broadcast later on.
+        if values.ndim == 1 and any(ax == 1 for ax in cond.shape):
+            cond = cond.ravel()
+
         # our where function
         def func(cond, values, other):
             if cond.ravel().all():
@@ -2817,10 +2822,10 @@ def _can_hold_element(self, element):
 
     def _try_coerce_result(self, result):
         if (isinstance(result, np.ndarray) and
-                np.ndim(result) > 0
-                and not is_sparse(result)):
+                np.ndim(result) == 1 and
+                not is_sparse(result)):
             result = SparseArray(result, kind=self.kind,
-                                 fill_value=self.fill_value, dtype=self.dtype)
+                                 fill_value=self.fill_value)
         return result
 
     def __len__(self):