mtmd : use align_corners for qwen3vl vision position embedding interpolation (#25781)

mtmd : use align_corners for qwen3vl vision position embedding interpolation (#25781)

The Qwen3-VL learned position embedding is interpolated to the runtime patch
grid with the default bilinear+antialias (align_corners=False) sampling, while
the transformers reference uses align_corners=True (torch.linspace(0, side-1, T)).
The mismatch scales grounding coordinates about the image center, growing with
image size and per-axis for non-square images (see #16880).