Store ConvTranspose2d weights in MLX (out, kh, kw, in) layout

Relayout of the 12 `*.resamplers.N.0.weight` tensors from torch (in, out, 2, 2) to MLX (out, 2, 2, in) so mlx-vlm loads them directly into nn.ConvTranspose2d without transposing at load time. All other tensors are byte-identical; model outputs are unchanged.

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:46b886fb6fca6dccdc5338a67225bf6727efe480700d705f0c2ac6601ef12185
-size 1481203814
+oid sha256:d21ea545b5f760b0d34d6538902378f6daa9d4c79dc5843891bbebec7c4da87b
+size 1481204170