Revert "Use random64 in Fischer-Yates algorithm for large N (#143682)"

This reverts commit 7013be0094e8d3ded2ba2f948082f98d63e622bb.

Reverted https://github.com/pytorch/pytorch/pull/143682 on behalf of https://github.com/wdvr due to failing Meta internal tests that need to be updated ([comment](https://github.com/pytorch/pytorch/pull/143682#issuecomment-2563487675))

test/test_tensor_creation_ops.py

@@ -3594,29 +3594,6 @@ class TestRandomTensorCreation(TestCase):
             self.assertEqual(non_contiguous_tensor, res)
             self.assertEqual(res.sort().values.long(), torch.arange(n, device=device))
 
-
-    @largeTensorTest("10GB", "cpu")
-    @largeTensorTest("40GB", "cuda")
-    @slowTest
-    def test_randperm_large(self, device):
-        # Test even distribution where rand32 might produce skewed "uniform" distribution
-        # n_items is chosen to not evenly divide 2**32 and be sufficiently large
-        # to easily detect skew
-        def decile(index, collection_size):
-            return index // (collection_size // 10)
-
-        n_items = 700_000_000
-        shuffled = torch.randperm(n_items, device=device)
-        interval = 1_000_000
-        shuffled_interval = shuffled[:interval]
-        # histogram implemented for float only
-        deciles = decile(shuffled_interval, shuffled.shape[0]).float().cpu()
-        hist, _ = deciles.histogram(10, range=(0, 10))
-        expected_bin = shuffled_interval.shape[0] / 10
-        expected_error = math.sqrt(expected_bin) / expected_bin * 3
-        error = (hist - expected_bin).abs().max() / expected_bin
-        self.assertTrue(error < expected_error, f"error {error} > {expected_error}")
-
     # Test exceptions when device and generator types are incompatible
     @onlyCUDA
     @unittest.skipIf(IS_FBCODE or IS_SANDCASTLE, "Produces inconsistent errors when run in fbcode.")