llama: add attn temperature tuning for llama arch (non-iswa)

[ngxson]

Add support for build_inp_attn_scale() in llm_build_llama

CC @ggerganov

[CISC]

Hmmm, I think this is basically the same as in Grok-2, except there the scale is calculated from temperature_length:
https://github.com/sgl-project/sglang/blob/86d10d220f665092f93a3f6e8a31a65a36a4f376/python/sglang/srt/layers/attention/triton_ops/decode_attention.py#L307-L311

Never implemented it (but the metadata is there as you can see) because I was unsure exactly where it should go.