server: concurrent router loads can leave one model queued forever #36

Upstream: ggml-org#28443

With router mode and --models-max 1, two simultaneous requests for unloaded models can leave one request queued indefinitely.

Environment: current main 465e49b9, Linux aarch64, CUDA. Three presets A/B/C used the same Gemma 4 model file to isolate scheduler behavior.

Reproduction:

1. Request A and let it become idle.
2. Request B and C simultaneously.
3. Apply a 25-second client timeout.

Observed: B completed in 3.06 seconds; C timed out after 25 seconds with zero bytes. Both waiters selected and attempted to unload A, but only B claimed the freed slot and C remained queued.

The affected LRU scheduler path was introduced in dff15d4a.