common : retry HTTP requests over IPv4 when IPv6 connect fails

[CptTZ]

Overview

On a host (e.g. HF) that has an IPv6 (AAAA) DNS record but no working IPv6 route, llama-server -hf hangs at the first HTTP request. With -v it stalls:

tonyz@tonyz-sff-lin ~/s/llama.cpp (master)> ./build/bin/llama-server -hf xxx -c 0 -v
0.00.083.740 I common_params_handle_remote_preset: looking for remote preset at https://huggingface.co/xxx/resolve/main/preset.ini
0.00.084.861 D common_download_file_single_online: no previous model file found xxx

Upon investigation, hang is in cli.Head(). huggingface.co has multiple addresses and cpp-httplib tries them sequentially with IPv6 first, and it never falls back to the working IPv4 address. The default connection timeout at 300s, the request blocks for ~5 minutes before failing.

This PR bounds the connect timeout and adds an IPv4 fallback, so dual-stack hosts with broken IPv6 connectivity work again -

tonyz@tonyz-sff-lin ~/s/llama.cpp (master)> ./build/bin/llama-server -hf xxx -c 0 -v --no-mmap
0.00.092.428 I common_params_handle_remote_preset: looking for remote preset at https://huggingface.co/xxx/resolve/main/preset.ini
0.00.093.583 D common_download_file_single_online: no previous model file found xxx
0.05.295.020 I common_download_file_single_online: HEAD failed, status: 404
0.05.296.016 I common_params_handle_remote_preset: no remote preset found, skipping
0.15.671.972 D common_download_file_single_online: using cached file: /home/tonyz/.cache/huggingface/hub/xxx.gguf
0.15.672.141 D common_download_file_single_online: using cached file: /home/tonyz/.cache/huggingface/hub/xxxx.gguf
0.15.672.533 I common_params_print_info: build 9439 (55f5dd2cf) with GNU 13.3.0 for Linux x86_64

Notes

• Chose cpp template for easier adaption of different httplib methods/signatures
• Timeouts: I picked 5 seconds connect timeout without a specific reason (libcurl defaults to 300ms, happy to use the same setup if needed), and it only bounds connection setup, not transfers, so it's safe.

Additional information

In #18828 we removed libcurl and in previous changes replaced libcurl with cpp-httplib - this caused a regression as libcurl supports rfc6555 - this is not supported in cpp-httplib, thus this issue.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, assisted with root-causing hang llama-server start.