sd: sync with master-849-d04e895

[wbruna]

master-849-d04e895:

• feat: prefetch streamed layers during compute
• refactor: unify runner lifecycles and weight residency
• feat: add verbose logging and log-level selection
• feat: enable single-GPU auto-fit with tiered parameter placement
• fix: reuse graph cut plans across CFG passes
• refactor: split ggml extensions and move implementations to cpp files
• fix: preserve K-quantized embedding weights
• fix: correct SDXL embeddings loading

We already got a lot of big refactors. On our side, beside normal merge conflicts:

• moved some of the KCPP_SD_MAINLINE_ tests due to the ggml_extend.hpp split;
• cleaned up the logging divergence, adding a bit of logic so the sd.cpp with our changes still keep its logging as-is;
• removed some LTX audio VAE parameters related to weight cleanup;
• dropped stream_layers parameter

Only lightly tested for now. I'm considering merging the next releases first, then testing all at once.

This is on top of #2452 . Let me know if I should split this up further.