Add targets_mv / frame_center_mv / corr_offset_mv overloads to
correlatePairs() and correlatePair().  When targets_mv is non-null,
GpuQuad.setRectilinearMovingTasks() generates per-tile TpTask centers
that follow the target at each frame (cx0 + vx*offset_scale), so the
GPU correlates tiles centered on the moving object and the result is the
residual velocity rather than the full VX displacement.

Callers that don't need this pass through the old uniform-grid overload
unchanged.  Also reverts the wrong "fixed-destination / shifted-read"
masking edit from the previous session; the masking was already correct
(mask tracks the target per frame) before that edit.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>