`Tensor.mean` backward fails when `axis` is a tuple

[github-actions]

File: leanpass/tensor.py

The backward implementation uses np.expand_dims(grad, axis=axis), but np.expand_dims accepts a single integer axis, not a tuple. When axis is a tuple (e.g., axis=(1,2)), this raises a TypeError. The code should instead reshape the gradient to keep the reduced dimensions and then broadcast.

Suggested fix:

if isinstance(axis, tuple):
+    for ax in sorted(axis, reverse=True):
+        grad = np.expand_dims(grad, axis=ax)
else:
+    grad = np.expand_dims(grad, axis=axis)

Filed automatically by ai-issue-scan.