fix: OTA target scaling uses height for both dimensions

[Mr-Neutr0n]

Bug

OTA target coordinate normalization uses the height dimension for both x and y scaling. The x coordinate should be normalized by width and y by height.

Fix

Corrected the scaling to use the appropriate dimension for each coordinate.

In utils/loss.py, the target xywh coordinates are converted from normalized [0,1] space to pixel space. The original code multiplies all four values (x, y, w, h) by imgs[batch_idx].shape[1] (height), but x and w should use shape[2] (width) instead. This affects ComputeLossOTA, ComputeLossBinOTA, and ComputeLossAuxOTA.

The fix changes the scalar multiplication to a per-coordinate tensor [W, H, W, H] so that x/w scale by width and y/h scale by height.