Preserve per-frame intrinsics in camera conditioning

[zjn20030811]

Summary

• preserve frame-varying fx, fy, cx, and cy instead of repeating the first calibration across every latent frame
• resample intrinsics on the exact same target timeline used for pose interpolation in both standard and fast inference
• keep singleton calibration and action-to-camera trajectories on the existing static-intrinsics path
• validate short/malformed calibration sequences and retain dtype/device, including bfloat16 inputs

Motivation

The documented camera-control input contains one [fx, fy, cx, cy] row per frame. Both inference paths currently reduce that array to Ks[0] before constructing ray/Plücker conditioning, so zoom or other dynamic calibration is lost while the pose remains time-varying.

The included examples/01 sequence has 561 per-frame rows and roughly a 29-pixel focal-length range. Repeating its first row changes the resulting edge rays by up to about 1.52 degrees. This patch aligns calibration and pose on one shared source/target grid before ray construction.

Compatibility

• a single calibration row is still repeated
• action-to-camera trajectories still use the first row because their generated pose timeline does not correspond to the external calibration timeline
• extra calibration rows are truncated with the pose sequence; a non-singleton short sequence now raises a clear error

Validation

• python -m pytest tests/test_cam_utils.py -q — 13 passed
• real example-array smoke checks for standard (41 frames) and fast (39 frames) pose/K/Plücker alignment
• full Ruff check for the new tests, fatal-rule checks for changed source files, py_compile, and git diff --check