server: multi-choice completions report usage for only one choice #55

Upstream: ggml-org#29451

With n > 1, chat completions return all choices but count only one choice in usage.completion_tokens.

Confirmed on Spark Linux ARM64, CPU Release, upstream 95887577ab5fead779581a7030a83c7752ff3234, Qwen2.5-0.5B-Instruct Q4_K_M.

llama-server -m qwen2.5-0.5b-instruct-q4_k_m.gguf -ngl 0 -c 2048 -np 4 -t 2 --host 127.0.0.1 --port 18927
curl http://127.0.0.1:18927/v1/chat/completions -H 'Content-Type: application/json' -d '{"model":"x","messages":[{"role":"user","content":"Write a long, detailed paragraph about the ocean and its many creatures."}],"max_tokens":8,"temperature":0.8,"seed":1,"n":4}'

Observed: four distinct choices, all finish_reason:length; server logs confirm 8 generated tokens each (32 total). Response reports completion_tokens:8, prompt_tokens:43, total_tokens:51. Controls with n=1/2/4 all report completion_tokens=8.

Cause: non-streaming result aggregation appends choices into arr[0] without summing usage. Introducing commit not bisected.