Use gesvd solver for SVDQuant

```python
# CUDA：只能用gesvd，保证不报错、精度不丢
if original_device.type == "cuda":
    u, s, vt = torch.linalg.svd(weight_f64, driver="gesvd", full_matrices=False)
# CPU：用默认gesvd，速度最快
else:
    u, s, vt = torch.linalg.svd(weight_f64, full_matrices=False)
```