Use torch.median for error stats (#16683)

### Summary
This is a re-land of #16673
with updated test data. Previously, the error stats logic would throw
when a large tensor is passed in due to size limitations of
torch.quantile. The above PR switched it to use torch.median, but this
has differing behavior for tensors with an even number of elements.

Quantile will return the mean of the two medians, but torch.median takes
the lower. Either is fine for this use case - we just need to update the
test logic accordingly. I verified that the error stats unit tests pass
with this change.

Author: GregoryComer

backends/test/harness/error_statistics.py

@@ -33,7 +33,7 @@ def from_tensor(cls, tensor: torch.Tensor) -> "TensorStatistics":
         return cls(
             shape=tensor.shape,
             numel=tensor.numel(),
-            median=torch.quantile(flattened, q=0.5).item(),
+            median=torch.median(flattened).item(),
             mean=flattened.mean().item(),
             max=flattened.max().item(),
             min=flattened.min().item(),

backends/test/harness/tests/test_error_statistics.py

@@ -14,7 +14,9 @@ def test_error_stats_simple(self):
         # Check actual tensor statistics
         self.assertEqual(error_stats.actual_stats.shape, torch.Size([4]))
         self.assertEqual(error_stats.actual_stats.numel, 4)
-        self.assertEqual(error_stats.actual_stats.median, 2.5)
+        self.assertEqual(
+            error_stats.actual_stats.median, 2
+        )  # torch.median takes the lower median
         self.assertEqual(error_stats.actual_stats.mean, 2.5)
         self.assertEqual(error_stats.actual_stats.max, 4)
         self.assertEqual(error_stats.actual_stats.min, 1)
@@ -44,15 +46,15 @@ def test_error_stats_different_shapes(self):
         # Check actual tensor statistics
         self.assertEqual(error_stats.actual_stats.shape, torch.Size([4]))
         self.assertEqual(error_stats.actual_stats.numel, 4)
-        self.assertEqual(error_stats.actual_stats.median, 2.5)
+        self.assertEqual(error_stats.actual_stats.median, 2)
         self.assertEqual(error_stats.actual_stats.mean, 2.5)
         self.assertEqual(error_stats.actual_stats.max, 4)
         self.assertEqual(error_stats.actual_stats.min, 1)
 
         # Check reference tensor statistics
         self.assertEqual(error_stats.reference_stats.shape, torch.Size([2, 2]))
         self.assertEqual(error_stats.reference_stats.numel, 4)
-        self.assertEqual(error_stats.reference_stats.median, 3.5)
+        self.assertEqual(error_stats.reference_stats.median, 3)
         self.assertEqual(error_stats.reference_stats.mean, 3.5)
         self.assertEqual(error_stats.reference_stats.max, 5)
         self.assertEqual(error_stats.reference_stats.min, 2)