Skip extracting meta tensor info for sparse tensors (#676)

* Skip extracting meta tensor info for sparse tensors

* Change expected failures

Author: anijain2305

functorch/_src/python_key.py

@@ -57,7 +57,8 @@ def __new__(cls, elem, proxy):
 
         r = torch.Tensor._make_subclass(cls, elem, elem.requires_grad)
         r.proxy = proxy
-        proxy.node.meta['tensor_meta'] = _extract_tensor_metadata(r)
+        if not elem.is_sparse:
+            proxy.node.meta['tensor_meta'] = _extract_tensor_metadata(r)
         return r
 
     def __repr__(self):

test/test_pythonkey.py

@@ -190,7 +190,6 @@ def f(x):
 
 
 make_fx_failures = {
-    xfail('to_sparse'),
     xfail('allclose'),
     xfail('nn.functional.dropout'),
     xfail('linalg.eigvals'),