fix(sdk): use resolved device for VLM encoder

[RemiliaForever (RemiliaForever)]

Summary

• Use the resolved llama.cpp device selection when initializing the VLM vision encoder.
• Disable mtmd GPU selection by default so CPU and hybrid paths remain on CPU.
• Preserve CPU fallback when the selected accelerator cannot initialize.

Test plan

• clang-format-18 on sdk/plugins/llama_cpp/src/vlm.cpp
• QCS Windows ARM64: cpu → CPU CLIP; gpu → OpenCL CLIP; npu → HTP0 CLIP; hybrid → CPU CLIP; all VLM generations passed.
• QCS9075M Linux: cross-compiled SDK successfully; cpu and hybrid -ngl 0 VLM cells passed, with logs confirming CPU CLIP.
• QCS9075M Linux gpu: resolved GPUOpenCL and CLIP selected OpenCL, but the platform Adreno driver aborts with back-end instruction validation failed during kernel compilation.
• QCS9075M Linux npu: CLIP selected HTP0, but only work with small model.
• git diff --check

Tracks qcom-ai-hub/geniex#1562