feat: MetalRT VLM backend + bolder screen overlay

[AmanSwar]

Summary

• MetalRT VLM backend: VLM commands (vlm, camera, screen) now use MetalRT's native vision pipeline when running on MetalRT engine. Falls back to llama.cpp gracefully if no MetalRT VLM model is found in HF cache.
• Screen capture overlay: Bolder border (8px), larger corner handles (28px), wider edge grab zones (20px), added edge midpoint handles, double-layer glow, and heavier label font.

Changes

• New MetalRTVlmEngine class (metalrt_vlm_engine.h/.cpp) wrapping MetalRT vision C API via dlsym
• vlm_init_locked() tries MetalRT first, falls back to llama.cpp (no longer hard-rejects MetalRT backend)
• All VLM functions (rcli_vlm_analyze, rcli_vlm_analyze_stream, rcli_vlm_get_stats, rcli_vlm_exit, handle_screen_intent) branch on vlm_use_metalrt flag
• Updated error messages to be backend-agnostic
• rcli_overlay.m visual improvements for easier drag/resize

Test plan

• Run rcli vlm <image> "describe this" on MetalRT engine — should use MetalRT VLM
• Run rcli vlm <image> on llama.cpp engine — should use llama.cpp VLM as before
• Run rcli vlm on MetalRT without VLM model in HF cache — should fall back to llama.cpp
• Run rcli screen — verify overlay is bolder with larger handles
• Verify overlay drag and resize from corners and edges