feat: Add Self-Healing Llama example with AVD and EDCC architecture

[iamrealvinnu]

Summary

This PR introduces a standalone example of Asynchronous Self-Healing Inference for Llama models. It demonstrates how to exploit Apple Silicon's Unified Memory Architecture to run a background verification process on the Neural Engine (ANE) while the GPU handles token generation.

Technical Innovation

1. Asynchronous Verification Daemon (AVD): Offloads hallucination detection to the ANE (via Core ML), ensuring zero overhead on the primary generation thread.
2. Head-Specific Causal Pruning: Injects 4D Gaussian masks into the attention mechanism to surgically excise logical drift from reasoning heads while preserving linguistic fluidity.
3. Entropy-Driven Context Compaction (EDCC): Physically deallocates low-entropy/masked nodes from physical RAM during natural pauses, enabling effectively infinite context loops.

Value to MLX Ecosystem

This is the first example in the repository demonstrating:

• Parallel utilization of ANE and GPU on a shared memory manifold.
• Dynamic, in-place KV cache mutation using MLX advanced indexing.
• Hardware-isolated runtime governance for high-stakes reasoning.

Setup & Usage

Instructions are provided in the sub-directory README.md. Includes a pre-compiled mock ANE model for instant verification.