Add decoder inference Colab (segmentation, depth, normals)

[bingyic]

Summary

Adds a new Colab notebook demonstrating inference with all three TIPSv2 DPT decoder heads:

1. Semantic Segmentation — ADE20K (150 classes) with color palette visualization
2. Monocular Depth Estimation — NYU Depth V2 samples with turbo colormap
3. Surface Normal Estimation — NYU samples with RGB normal map visualization

Notebook Structure

Section	Description
Setup	Install deps, clone repo
Download	Checkpoints (vision encoder + 3 DPT heads) + sample images (ADE20K, NYU)
Load Models	Vision encoder + SegmentationDecoder, DepthDecoder, NormalsDecoder
Inference	Segmentation on ADE20K, Depth + Normals on NYU (3 sample images each)

Notes

• Supports all variants: B, L, So, g (selectable via dropdown)
• GPU runtime recommended for faster inference
• Uses bicubic upsampling + L2 re-normalization for smooth surface normal maps