Sync SD model and LLAMA model loading / unloading

[fszontagh]

• if a llama model is required to load, unload SD model from RAM/VRAM
• if a sd model is required to load, unload LLAMA model from RAM / VRAM

[iwr-redmond]

Please consider adding an override option similar to the existing one so that users with higher-VRAM cards can opt to keep both in memory.

For example, I reckon that a 1.5B Q4_0 LLM (~1.5GB) and an SD1.5 checkpoint (~2GB) would be able to coexist on an 8GB video card. This should allow, for example, the NexaQuant Qwen R1 model to run alongside Dreamshaper.