Fit estimates don't account for Unsloth Studio's beyond-VRAM offload

llmfit's compatibility scoring assumes standard offload paths (llama.cpp / Ollama / MLX-style CPU+GPU spill). Unsloth Studio uses dynamic quantization plus selective offload, which lets models run meaningfully larger than what fits in raw VRAM. llmfit does not currently model this, so it can under-predict what is actually runnable on a given machine.

Example: llmfit flags a ~27B model (e.g. Qwen3-27B) as not fitting a given laptop's VRAM, but the same model runs fine under Unsloth Studio on that hardware.

Ask: Add Unsloth Studio as a distinct backend/runtime profile with its own headroom calculation, separate from the generic CPU+GPU spill model, so fit estimates reflect its actual ceiling instead of the conservative VRAM-only bound.