[misc] allow ATTN_QAT_TRAIN on Wan cross-attention (train/infer parity)

The cross-attn (attn2) LocalAttention allowlist had ATTN_QAT_INFER (added in
hao-ai-lab#1598) but not ATTN_QAT_TRAIN, so QAT finetune silently ran cross-attn in Flash
while self-attn fake-quantized — a train/inference mismatch vs the FP4 inference
path (which quantizes both). Add ATTN_QAT_TRAIN so the finetune fake-quantizes
cross-attn too. (self-attn already had it via the DiT config list.)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: Mister-Raggs

fastvideo/models/dits/wanvideo.py

@@ -186,7 +186,8 @@ def __init__(self,
             causal=False,
             supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN,
                                           AttentionBackendEnum.TORCH_SDPA,
-                                          AttentionBackendEnum.ATTN_QAT_INFER))
+                                          AttentionBackendEnum.ATTN_QAT_INFER,
+                                          AttentionBackendEnum.ATTN_QAT_TRAIN))
 
     def forward(self, x: torch.Tensor, context: torch.Tensor,
                 context_lens: int):