Switch from inverse to linalg_inv_ex to avoid sync

[matthewdcong]

torch::inverse, an alias for torch::linalg_inv, synchronizes the host with the device after computing the inverse in order to inspect error codes. To avoid this, we switch to torch::linalg_inv_ex (as suggested by the PyTorch documentation) which exposes a check_errors=False parameter. This lets us avoid the synchronization by default but also enable it for debugging if needed.