fix(server): avoid phantom router entry on proxy last-used bump

[marksverdhei]

What

Fixes #154 by changing server_models::proxy_request to update last_used via mapping.find(name) instead of mapping[name]. If the model entry disappears during a concurrent reload/remove race, the bump is skipped instead of default-inserting an empty instance_t.

Why

proxy_request resolves metadata before proxying, then optionally bumps LRU state under the mutex. The old operator[] access could create a phantom UNLOADED model entry if the canonical model was erased between those steps. That phantom could appear in /models and later drive an empty-preset child spawn.

Validation

• git diff --check
• cmake --build build --target llama-server -j 8
• ./build/bin/llama-server --version