Add Apple Foundation Models provider support

[vincent-herlemont]

Apple recently released the Foundation Models framework (macOS/iOS 26+), enabling on-device LLM inference with no network dependency. This would be a valuable addition to Rig's provider ecosystem, especially for privacy-focused and offline applications on Apple platforms.

Unlike existing HTTP-based providers, Foundation Models requires native platform integration. The libai library provides C bindings that could bridge the Swift framework to Rust via FFI.

Key points:

• On-device inference, no API key required
• Platform-specific: macOS 26.0+ / iOS 26.0+ / visionOS
• 4096 token context limit
• Supports streaming responses and tool calling
• No embeddings API available (completion only)
• Would require #[cfg(target_os = "macos")] feature gating

Open questions:

• Should this live in-tree or as a community crate?
• How to handle availability checks (SystemLanguageModel.isAvailable)?
• Testing strategy given platform requirements

References:

• Apple Foundation Models documentation
• Apple Machine Learning overview
• libai - C bindings for Foundation Models
• Rig provider implementation guide
• Rig provider clients concepts

Happy to contribute an initial implementation if there's interest.

[joshua-mo-143]

Looks good. We should add this as a companion crate though, I have a feeling that this will require adding a lot of new crates.