Ollama copies GGUF files, lacks in-place mode

Ollama's simplicity is a feature until you care about storage semantics. For local benchmarking, the missing zero-copy path is a real papercut because it adds disk overhead without changing the model itself.

• –The official import docs support `FROM /path/to/file.gguf` plus `ollama create`, but they do not describe a first-class in-place serving mode.
• –An open GitHub issue says `ollama create` currently does a regular copy of the `.gguf` file and asks for copy-on-write or reflink behavior instead.
• –For teams comparing Ollama with llama.cpp, that means duplicate model storage becomes part of the workflow even if runtime throughput is the only metric under test.
• –The issue points to APFS, Btrfs, and ZFS as the kinds of filesystems where reflinks could help, which shows the current copy behavior is an implementation choice, not a GGUF limitation.