Add local Qwen3.5 vision functions with ordered text and image inputs - #12

The Python SDK previously accepted only text inputs. This adds local Qwen3.5-0.8B image functions that accept ordered positional strings and paw.Image snapshots, through both prompted adapter-free functions and compatible compiled .paw bundles. Existing text function calls keep their interface.

Images are converted to native-size RGB PNG with transparency composited over white; patch sizing belongs to the backend. Versioned manifests pin model/projector assets and preprocessing. Matching live functions share a serialized runtime, switch LoRAs, and clear recurrent state between calls. The optional vision extra supplies Pillow and requires llama-cpp-python 0.3.35–0.3.36.

Image-runtime calls keep returning strings by default. Passing return_info=True returns an immutable FunctionResult with text, backend finish reason/token counts, and elapsed call time including preparation, lock waits, inference and cleanup. Missing backend metadata remains None; failures still raise.

Validation:

• 621 hermetic tests passed on Python 3.12 and 3.14; 578 passed with one expected release-tooling skip on Python 3.8.
• The underlying vision runtime passed 11/11 real offline inference checks on each of llama-cpp-python 0.3.35 and 0.3.36 on macOS Metal, including image ordering, prompted/LoRA interleaving and shared-runtime lifecycle. The added metadata path is covered by hermetic native-boundary tests, including concurrent-call isolation, truncation, timing, zero-token calls and error propagation.
• Four documentation examples passed with a mocked backend. Built wheel/source contents were allowlist-checked and matched to the final source.

Real vision inference on Linux/CUDA and Windows has not been qualified here. Hosted image compilation, image export workflows and remote image inference remain outside this PR.