Migrate Ideogram runtime to MLX

[haxlys]

Summary

• Replace the legacy PyTorch/MPS runtime path with an MLX/mflux runtime targeting MLXBits/ideogram-4-mlx-q8.
• Add server/mlx_runtime.py and ideogram4_mlx.py; remove old MPS benchmark/LoRA helper scripts.
• Keep FastAPI/WebUI contracts stable while surfacing MLX backend, model repo, quantization, memory, and local LoRA reload state.
• Route all MLX/mflux model operations through a single worker thread to avoid thread-local MLX stream errors during LoRA reload + generation.
• Add ./run.sh doctor and benchmark documentation.

Validation

• python -m compileall server ideogram4_mlx.py scripts/doctor.py
• rg "torch|safetensors.torch|from ideogram4|import ideogram4" server ideogram4_mlx.py
• ./run.sh doctor (0 failures; 1 warning for unset optional IDEOGRAM4_MLX_CACHE_LIMIT_GB)
• cd webui && pnpm lint
• cd webui && pnpm build
• Full local smoke: Magic Prompt health/generation, model load, 256x256 V4_TURBO_12 generation, LoRA apply -> generate -> remove -> generate.

Operational Notes

• legacy/pytorch-mps has been pushed separately to preserve the old PyTorch/MPS runtime as the rollback branch.
• IDEOGRAM4_MODEL_DAEMON_AUTOLOAD now defaults to 0 to avoid immediately reserving about 29GB of unified memory while a local Magic Prompt LLM may also be running.
• Use the WebUI Load button or POST /api/model/load when image generation is needed.
• IDEOGRAM4_MLX_CACHE_LIMIT_GB remains optional for machines that need a stricter reusable MLX cache budget.

Follow-up Tracking

• mflux remains pinned to PR #445 commit 8d80b9cb53688b62a2f814604b9f8b48987c5acd until MLXBits q8 loading lands in a stable release.
• Stable-release migration is tracked in Track stable mflux release for MLXBits Ideogram q8 loader #2.