fix(wan): WanModel.patchify flatten + return grid size tuple

Fixes #1063.

WanModel.forward (and the parallel usp_dit_forward in
diffsynth/utils/xfuser/xdit_context_parallel.py) both unpack
patchify's return value as `x, (f, h, w) = self.patchify(x)`, but
patchify only returns `x` -- and that `x` is a 5D (B, dim, f, h, w)
tensor straight out of Conv3d, not the (B, f*h*w, dim) sequence the
transformer blocks expect.

Net result: every Wan training run crashes at the first forward call
with `ValueError: not enough values to unpack (expected 2, got 1)`.
The block loop never executes.

The fix: extract (f, h, w) from the Conv3d output shape, flatten the
3D spatial-temporal grid into a token sequence (B, f*h*w, dim) ready
for the transformer blocks, return (x, (f, h, w)) to match what
forward and unpatchify expect.

Verified end-to-end with a synthetic 8-layer WanModel on the same
forward path (input shape (1, 16, 4, 8, 8) -> output shape
(1, 16, 4, 8, 8) round-trip), forward + backward both complete,
loss gradient propagates through all blocks.

Author: genno-whittlery

diffsynth/models/wan_video_dit.py

@@ -498,7 +498,13 @@ def patchify(self, x: torch.Tensor, control_camera_latents_input: Optional[torch
             y_camera = self.control_adapter(control_camera_latents_input)
             x = [u + v for u, v in zip(x, y_camera)]
             x = x[0].unsqueeze(0)
-        return x
+        # After Conv3d the tensor is (B, dim, f, h, w). The forward expects
+        # (x, (f, h, w)) where x has been flattened to (B, f*h*w, dim) ready
+        # for the transformer blocks; the grid size tuple drives the RoPE
+        # frequency cat and the unpatchify rearrange at the end.
+        f, h, w = x.shape[-3], x.shape[-2], x.shape[-1]
+        x = x.flatten(2).transpose(1, 2)
+        return x, (f, h, w)
 
     def unpatchify(self, x: torch.Tensor, grid_size: torch.Tensor):
         return rearrange(