Check state reorder implementation in cross attention

We have a performance optimization in `StaticAttentionState` where we avoid reordering the KV cache if the number of beams has not changed (see [here](https://github.com/facebookresearch/fairseq2/blob/f2be5c0196ac9eb566a1d2bc5ef3544f59220a83/src/fairseq2/nn/transformer/multihead_attention.py#L888)). However in our beam search implementation, it is theoretically possible to have the same number of sequences in a batch even when the number of beams changes. Although a very edge case, it would be nice to ensure that we handle it correctly. Check whether we can address it without introducing too much complexity.