Fix Ministral3 and Mistral4 ignoring packed sequences under FlashAttention (#50450)
Ministral3Attention and Mistral4Attention take position_ids as a named
argument (for the Llama 4 query scaling) and no longer forwarded it to the
attention function, so FlashAttention could not detect padding-free packed
rows and let documents attend across boundaries. Forward position_ids to
attention_interface, as the other Mistral-family models do.