fix(server): expose loaded model metadata

[derogab]

Summary

Fixes model metadata reporting for /v1/models and /v1/models/<id>.

The server loads exactly one GGUF at startup, but previously model discovery advertised both deepseek-v4-flash and deepseek-v4-pro. This could mislead clients into thinking both variants were available at the same time.

This PR makes model metadata reflect the loaded GGUF:

• Flash GGUF exposes only deepseek-v4-flash
• Pro GGUF exposes only deepseek-v4-pro
• /v1/models/<id> returns 404 for the non-loaded model id

The request model field still does not select a different GGUF; inference uses the single loaded model.

Commits

• 77c27a5 fix(server): report pro model name in models endpoint
• 059d6cf fix(server): expose only loaded model metadata

Testing

Tested on Mac CPU build, 36 GB:

./ds4_test --server

  Result:

  server:
  server: OK
  ds4 tests: ok

Tested on NVIDIA DGX Spark / GB10, 128 GB:

make ds4_test
./ds4_test --server

  Result:

  server:
  server: OK
  ds4 tests: ok

Live CUDA server test on NVIDIA DGX Spark / GB10, 128 GB, with Flash GGUF:

./ds4-server --cuda --ctx 2048 --tokens 256 --host 127.0.0.1 --port 8000
curl -s http://127.0.0.1:8000/v1/models | jq -r '.data[] | "\(.id) => \(.name)"'

  Result:
  deepseek-v4-flash => DeepSeek V4 Flash

Loaded model endpoint:
curl -s http://127.0.0.1:8000/v1/models/deepseek-v4-flash | jq -r '.id + " => " + .name'

  Result:

  deepseek-v4-flash => DeepSeek V4 Flash

Non-loaded model endpoint:
curl -s -w '\nHTTP %{http_code}\n' http://127.0.0.1:8000/v1/models/deepseek-v4-pro

  Result:
  
  {"error":{"message":"unknown model","type":"invalid_request_error"}}
  
  HTTP 404