Handle parallel optical axes in focus-based camera centering

[Afloat16]

Focus-based camera centering currently inverts a 3x3 normal matrix. Parallel optical axes make that matrix rank-deficient, so a normal forward camera trajectory can abort dataset preparation with linalg.inv instead of producing a focus estimate.

Use f_prior + A^+ (b - A f_prior) to select the least-squares minimizer nearest the current focus estimate. Full-rank configurations retain the existing solution; parallel cameras preserve the unidentifiable depth along their optical axis. Match the projector identity's device/dtype to the input poses.

Validation:

• 9 focused tests passed, 1 CUDA test skipped on CPU Torch 2.14.1. Unchanged source: 6 failed, 3 passed, 1 skipped.
• Tests cover parallel float32/float64 axes, nullspace-prior preservation, actual auto_orient_and_center_poses dataset preparation, rigid-world equivariance, a noisy full-rank reference, active-camera fallbacks, and nearly parallel scale consistency.
• All existing cameras tests plus the new regressions: 18 passed, 1 skipped (existing deprecation warnings).
• Repository Ruff 0.12.2 configured lint/format checks on changed files, license check and git diff --check passed.

Full project ns-dev-test, GPU execution, and reconstruction-quality benchmarks were not run for this patch. Focus-estimator derivatives with respect to pose parameters are not claimed.