Fix Llama4 text models with `attention_chunk_size=None` (Llama Guard 4) (#50763)
* Fix Llama4 text models with attention_chunk_size=None
Llama Guard 4 sets attention_chunk_size to None, but the RoPE layers
were still typed as chunked_attention, so building the cache or the
chunked mask failed. Only mark layers as chunked when a chunk size is
set, and only build the chunked mask when a layer uses it.
* Build the chunked mask lazily like laguna; drop the standalone test