fix(server): recover stalled HF model downloads

[fl0rianr]

Fixes model downloads that can hang indefinitely at 0 bytes for Hugging Face /resolve/ URLs backed by Xet.

The downloader now detects no-byte progress stalls and retries affected Hugging Face resolve downloads with a byte-range request. Non-Hugging-Face downloads keep the existing behavior.

Fixes #2416

[The-Monk]

Verified on RDNA4 / gfx1201 (2× Radeon AI PRO R9700, lemonade-server 10.8.1 + this PR built from source).

Multiple Xet-backed repos that hung indefinitely at 0 bytes on stock 10.8.1 now pull cleanly with this patch:

repo (Xet-backed)	result
nomic-ai/nomic-embed-text-v1-GGUF	✅ 7s
unsloth/Qwen3-Embedding-0.6B-GGUF	✅ 14s (609 MB)
Qwen3.5-0.8B-FP16-vLLM (13-file safetensors)	✅ 42s (1.6 GB)

All confirmed Xet via the /resolve/ → cdn.hf.co/xet-bridge-us 302 + x-xet-hash header. The byte-range stall recovery does the job. 👍 (Also confirms the root cause: the env opt-outs HF_HUB_DISABLE_XET/HF_XET_DISABLE had no effect since lemond's downloader is native C++, not the Python huggingface_hub client.)