docs(mlx): add Apple Silicon execution backends section

[jasperf]

Summary

Adds an "Execution Backends on Apple Silicon (Metal / MLX / ANE)" section to docs/MLX-RUNTIME.md, answering a recurring conceptual question: whether there are models that run "on CPU" or better on Mac Silicon, as opposed to the RAM+GPU pattern.

The section clarifies:

• Unified memory means "RAM vs GPU" isn't a PC-style split — the GPU does the math on weights living in shared RAM.
• There's no separate class of "CPU model" — it's the same models run through a different backend.
• Compares the three backends in a table: Metal/GGUF (default, what the "spills X% to CPU" notes describe), MLX (Apple's unified-memory framework, the most optimized practical path), and the Neural Engine via Core ML (not viable for LLM generation today, due to the growing KV-cache of autoregressive decoding).
• Notes CPU-only inference only makes sense when weights don't fit the GPU ceiling.

Placed right after the Overview since it's foundational framing the rest of the doc builds on.

Changelog

Released as [1.6.2] — 2026-07-02.