[feat] WanTrack val: MotionStream-aligned joint text+motion CFG
Wire compositional text + motion CFG into TrackValidationCallback._sample,
following MotionStream Eq. 2 exactly:
v_no_text = v(∅, c_m) # drop text, keep tracks
v_no_motion = v(c_t, ∅) # keep text, drop tracks
v_base = α·v_no_text + (1-α)·v_no_motion, α = wt/(wt+wm)
v̂ = v_base + wt·(v_full - v_no_text) + wm·(v_full - v_no_motion)
3 NFE/step when either wt or wm != 1. At wt=wm=1 collapses to single v_full
(no CFG). No-track counterfactual branch keeps plain text CFG only (motion
CFG undefined without tracks). "Unconditional" text = zero embedding, matching
WANTRACK_TEXT_DROP training.
Configs updated to MotionStream's recommended wt=3.0, wm=1.5.
shaoxiongduan
aea5cadb5f