UPSTREAM PR #18054: server: Fix router proxying to child processes when --host is specified

[loci-dev]

Mirrored from ggml-org/llama.cpp#18054

llama-server --host now works for:

• specific external facing addresses both locally and externally (192.168.3.14)
• localhost (127.0.0.1)
• all interfaces (0.0.0.0)

It worked with 127.0.0.1 and 0.0.0.0 before, but when a specific external facing address was specified it:

• served the WebUI correctly on the specified ip address
• switch to 127.0.0.1 for child when model was loaded

[loci-review]

Explore the complete analysis inside the Version Insights

Based on the analysis results, I can now provide the summary:

---

Performance Analysis Summary

Condition 1 Applied: No meaningful changes detected

The comparison between version 55c2244c-43d3-4f65-8b57-f5d9ba63cda9 and baseline 24dfe6da-2760-403d-a4ec-e8afe1ec7d4f shows no measurable performance differences. Power consumption across all binaries remains unchanged, with negligible variations under 0.001% in libllama.so, llama-cvector-generator, llama-run, and llama-tts. No function-level performance regressions detected. Inference throughput and tokens per second remain unaffected.