fix: return correct tuple size from predict_with_cfg zero-init early-exit

[sajeel-haider]

Summary

• Fixes the ValueError: not enough values to unpack (expected 3, got 2) crash when both cfg_zero_star and use_zero_init are enabled in WanVideoExperimentalArgs
• Passes through cache_state/teacache_state unchanged in the zero-init early-exit path — at step 0 there is no cache update, so returning the input state is the correct identity behavior
• Fixes the same bug in both nodes_sampler.py (3-tuple callers) and skyreels/nodes.py (2-tuple callers where None was returned instead of state)

Changes

nodes_sampler.py (line 1185):

 if use_cfg_zero_star and (idx <= zero_star_steps) and use_zero_init:
-    return z*0, None
+    return z*0, None, cache_state

skyreels/nodes.py (line 460):

 if use_cfg_zero_star and (idx <= zero_star_steps) and use_zero_init:
-    return latent_model_input*0, None
+    return latent_model_input*0, teacache_state

Test plan

• Verified the fix matches the return signature of all other return paths in both functions
• Confirmed all callers in nodes_sampler.py unpack 3 values and all callers in skyreels/nodes.py unpack 2 values
• Test with Wan 2.2 I2V/T2V workflow with cfg_zero_star=True + use_zero_init=True

Fixes #2010