docs: restructure tutorials around runtime quantization concept

- Reframe Lighting as "runtime quantization" engine throughout all docs
- Reorder tutorials: 1) runtime quantization, 2) export, 3) download pre-exported models
- Rename tutorial files to match new numbering (tutorial-2-export, tutorial-3-download)
- Clarify export = persisting all runtime-quantized models to disk
- Replace "baked/烘焙" with "fused/融合" for LoRA terminology
- Fix SVDQ vs Lighting performance comparison (similar speed, Lighting ~20% faster on sub-RTX 50)
- Add workflow overview images and workflow file links to each tutorial overview
- Remove redundant "Basic Flow" section from READMEs
- Describe downloaded models as pre-exported by QuantFunc from runtime quantization