Fix llama-server version selection

[Rader]

What changed

• Prefer a llama-server installed next to the running csghub-lite binary before searching PATH.
• Cover the co-located binary path on macOS, Linux, and Windows.
• Replace existing shared-library compatibility links when the installer upgrades llama.cpp.

Why

Background services can receive a different PATH from the interactive shell used for installation. This allowed a stale system-wide llama.cpp build to be selected instead of the matching release installed beside csghub-lite. The installer could also leave compatibility symlinks pointing at libraries from an older release.

The mismatch caused newer multimodal embedding models to crash during llama-server startup even though the pinned llama.cpp release was already installed under ~/.local/bin.

Validation

• go test ./internal/inference ./internal/server
• sh -n scripts/install.sh
• git diff --check
• Verified both local Qwen3-VL embedding model IDs return HTTP 200 with 2048-dimensional text embeddings while using the co-located llama-server b9158.

[Rader]

fix embedding model run exception: #73