Fix streaming-inference alignment with training-time contracts

[study-overflow]

We hit this while evaluating LingBot-VA on streaming-inference mode: training loss was healthy and teacher-forced open-loop predictions were accurate, but closed-loop rollouts still collapsed with the same checkpoint — a symptom of inference-side train/test mismatch that compounds across chunks.

Fixes two streaming-inference bugs that decouple closed-loop rollout behavior from training.

1. Cold-start action conditioning now uses physical zero

   • At frame_st_id == 0, the server synthesizes an all-zero action history.
   • Old code built it directly in normalized space, so torch.zeros decoded to the per-channel statistical midpoint (q01 + q99) / 2 instead of a physical zero action.
   • Added _make_zero_action_condition() that starts from physical zero, applies the same quantile normalization as training, and masks unused channels.

2. KV-cache attention visibility now matches training mask

   • Training mask allows a noisy query to attend only to past clean tokens, same-frame noisy tokens, and past predicted tokens, all within the window.
   • Old inference surfaced every allocated cache slot, letting noisy queries leak same-frame clean answers and peek at future predicted tokens.
   • Added frame_id per slot, set_frame_context(), and _select_valid_cache_slots() to reproduce the training visibility rules exactly.

Validation:

• py_compile clean; git diff --check clean.
• Same checkpoint + env seed 100001 + inference seed 0 went from 600-step failure to 293-step success.

Files changed:

• wan_va/modules/model.py
• wan_va/wan_va_server.py