llama : fix n_rot default

[ggerganov]

fix #8246 #8251

The logic for determining default n_rot parameter did not take into account LLM_KV_ATTENTION_KEY_LENGTH overrides. This lead to invalid context shift for Gemma2 models:

# gemma-2-27b-it
n_embd = 4608
n_head = 32
n_embd / n_head = 144
n_embd_head_k   = 128   # (from KV header)

# old
n_rot = 144

# new
n_rot = 128

# gemma-2-9b-it
n_embd = 3584
n_head = 16
n_embd / n_head = 224
n_embd_head_k   = 256   # (from KV header)

# old
n_rot = 224

# new
n_rot = 256

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High