Fix world-to-camera transforms in flow keyframing

[zjn20030811]

Summary

• honor the world-to-camera extrinsic contract in both windowed streaming implementations
• unproject current-camera points with E_cur^-1 and reproject them with E_kf
• add an analytic projection regression with non-square geometry and anisotropic intrinsics

Root cause

pose_encoding_to_extri_intri returns OpenCV world-to-camera extrinsics, but the flow-keyframe code treated the current extrinsic as camera-to-world and inverted the keyframe extrinsic. It therefore evaluated

x_kf = E_kf^-1 E_cur x_cur

instead of the required relative transform

x_kf = E_kf E_cur^-1 x_cur

This affects both streaming and windowed keyframe selection paths in gct_stream_window.py and gct_stream_window_v2.py.

The regression uses a 24x40 image, fx != fy, different current/keyframe intrinsics, rotation, translation, and varying depth. The direct homogeneous-projection reference is 4.492928 px; both fixed production implementations return 4.492928 px, while the previous inverse convention returns 6.628602 px. This crosses the 5-pixel flow threshold and changes the keyframe decision.

Validation

• python -m pytest -q — 2 passed
• python -m compileall -q lingbot_map tests
• Black check for the new regression module
• Ruff fatal checks for the changed Python files
• git diff --check