fix VAE tiling for Qwen Image

[wbruna]

Intended for #851 , but works fine on current master. Edit: encoding path fix applied on top of qwen_image.

[Green-Sky]

Did you try if this works with wan?

[LostRuins]

It will not trigger with WAN in it's current state, because encode_video will be true, so it will never enter the vae tiling codepath

[LostRuins]

@wbruna i can confirm it works for qwen image, and there's no regression with WAN (still works fine with no tiling)

[wbruna]

It was failing simply because the split+merge assumed a specific tensor shape (and blowing up deep inside ggml because the wrong shapes triggered a division by zero). But for Wan, I'm sure leejet would have spotted this bug if it was just a matter of fixing the dimensions... Feel free to try it, though; I'm not sure I have either hardware or patience enough for that 😅.

[wbruna]

Not enough for img2img, unfortunately; any encode tiling triggers an assertion:

ggml/src/ggml-backend.cpp:300: GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds") failed

Edit: fixed it, but needed an explicit test for Qwen Image.