[LLM:Bugfix] Fix incorrect kv_seq_len in gen_attention_mask() for autoregressive decoding

[HirokenOvo]

During the decoding phase (auto-regressive generation), Llm::generate() calls forward(input_ids, is_prefill) with input_ids.size() == 1. The current logic in Llm::gen_attention_mask() sets kv_seq_len = seq_len (where seq_len=1 for decoding), which is incorrect. This ignores the accumulated history length（mContext->all_seq_len）, leading to mismatched attention masks and affecting the calculation of attention scores, thereby influencing the model's inference results.

[CLAassistant]

All committers have signed the CLA.