afm 0.9.7 adds Telegram, batch decoding

This looks less like a wrapper refresh and more like a serious local inference runtime for Apple Silicon Macs. The most interesting part is that the release is focused on throughput and reliability, not just model support.

• –Concurrent batch decoding and shared prefix caching should improve real-world throughput, especially for repeated prompts and multi-request workloads
• –XGrammar constraints for tool calls are a practical fix for brittle XML/JSON formatting issues on smaller or less compliant models
• –Telegram bridging makes the local model reachable from anywhere without exposing the machine directly to the public internet
• –The Swift-only stack plus Homebrew/PyPI install path keeps the barrier low for Mac developers who want local, OpenAI-compatible inference
• –This is still a niche, Mac-only layer, but it is a compelling one for Apple Silicon users who want private local AI with stronger API ergonomics