llama-server cache migration breaks GGUF scripts

Good ecosystem move, rough developer experience. If cache paths are part of how users deploy and share models, a one-way startup migration is a breaking change, not just a refactor.

• –The Hugging Face cache is blob/snapshot based, so direct filename assumptions stop working once files are re-homed.
• –Teams that symlink, rsync, or mount model directories across machines feel the pain first because the new layout optimizes dedupe, not human-readable paths.
• –`--model-url` users are spared, but anyone who leaned on `-hf` gets an abrupt filesystem change with no obvious rollback.
• –This should have shipped with a grace period, an opt-out, or a compatibility shim before moving user files.