Division by zero not guarded in Tensor.__truediv__ backward pass

[github-actions]

File: leanpass/tensor.py

The backward implementation computes -out.grad * self.data / (other.data ** 2). If other.data contains zeros, this yields infinities or crashes. NumPy will raise a runtime warning and propagate inf values.

Fix: add a small epsilon to the denominator or raise a clear error when dividing by a tensor that may contain zeros.

Label: bug

Filed automatically by ai-issue-scan.