Summary

Fix test_hymba

What Changed

1.shared_kv_cache_dict was only populated when reuse_kv=True.

That breaks models like Hymba where not every decoder layer has reuse_kv=True: earlier layers may need to publish KV for later layers even when the current layer itself does not consume kv_last_layer. In those cases,
prev_kv stays empty by the time a later layer actually needs it.

This change adds a model-level write_shared_kv_cache switch on BaseQModel, keeps the default behavior unchanged, and enables it for Hymba.

2.hymba is compatible with transformers v5.