Add DUSt3R stereo method

[nburrus]

Summary

• Adds stereodemo/method_dust3r.py — a minimal PyTorch inference adapter for DUSt3R (CVPR 2024), using a self-contained local port of the ViT-Large encoder + cross-attention decoder + DPT/linear head (no dependency on the official DUSt3R package)
• Adds stereodemo/dust3r_lib/ — minimal inference code adapted from ibaiGorordo/dust3r-pytorch-inference-minimal (CC BY-NC-SA 4.0)
• Supports the 512-dpt and 224-linear checkpoints (~2.3 GB each, downloaded on first use); non-commercial only
• Scale recovery: estimates DUSt3R's implicit focal length from the left point map and the predicted baseline from the right point map (rectified-pair assumption), then computes pixel disparity as f_dust3r × b_dust3r / Z_dust3r — avoids using the user's calibrated focal length in PnP, which would bias the scale

Test plan

• uv run stereodemo datasets → DUSt3R appears in the method list
• 512-dpt checkpoint downloads on first use and produces a plausible coloured point cloud
• uv run python tests/test_methods.py TestStereoInference.test_dust3r passes (skips if checkpoint not cached)
• No regression in other methods (test_bm, test_sgbm, test_crestereo, etc.)

🤖 Generated with Claude Code