PR #1745 merged: install from upstream mlx-vlm, retire the fork branch reference

Correct the activation figure: 14 GB is the 8192-token budget, not the 4096 default

Card: torch-free install note (discussion #1, fixed in branch bf9ca87)

Card: link mlx-vlm PR #1745; verified snippet, no trust_remote_code needed

Mage-VL 8-bit (mlx_vlm affine g64, 9.023 bpw) — 48/48 greedy-token parity with the PyTorch reference