Silent wrong CER computation may be going on in WerMetric

**Describe the bug:** WerMetric expects a Tensor as ref_seqs and hyp_seqs. However, editdistance silently gives the wrong result on tensors, see image: <img width="494" height="208" alt="Image" src="https://github.com/user-attachments/assets/e14d78d7-fc37-4c82-b88e-7fecc72b3cfa" /> **Describe how to reproduce:** We found that while debugging parity issues with the MMS OSS release. **Describe the expected behavior:** One solution is to call .cpu().tolist() **Environment:** - **Additional Context:** -