llama : remove obsolete n_pad from memory instances (#31200)
Every caller passes n_pad = 1 and the only two users clamp it to 256.
Drop the parameter, the members and the trivially true kv_size assert.
Assisted-by: dsh:llama.cpp/ggml-org/Qwen3.8-Flash-Next-GGUF (MacBook, M5 Max, 128GB)
