Fix regression in object detection models

[alanocallaghan]

Description

Resolve #3840

Previously imageWidth/height was used from ObjectDetectionTranslator:

djl/api/src/main/java/ai/djl/modality/cv/translator/SingleShotDetectionTranslator.java

Lines 63 to 66 in 3cc328d

	double x = imageWidth > 0 ? box[0] / imageWidth : box[0];
	double y = imageHeight > 0 ? box[1] / imageHeight : box[1];
	double w = imageWidth > 0 ? box[2] / imageWidth - x : box[2] - x;
	double h = imageHeight > 0 ? box[3] / imageHeight - y : box[3] - y;

These were usually ignored/set to zero unless the rescaleRatio was set:

djl/api/src/main/java/ai/djl/modality/cv/translator/ObjectDetectionTranslator.java

Lines 81 to 92 in 3cc328d

	/**
	* Sets the optional rescale size.
	*
	* @param imageWidth the width to rescale images to
	* @param imageHeight the height to rescale images to
	* @return this builder
	*/
	public T optRescaleSize(double imageWidth, double imageHeight) {
	this.imageWidth = imageWidth;
	this.imageHeight = imageHeight;
	return self();
	}

now, width/height are used from BaseImageTranslator, and these always default to 224x224:

djl/api/src/main/java/ai/djl/modality/cv/translator/BaseImageTranslator.java

Lines 111 to 112 in 3edd20a

	protected int width = 224;
	protected int height = 224;

Currently it is possible (and indeed common) to double-normalize the coordinates in this method (if width/height are nonzero and applyRatio is true), and the default is to normalize coordinates even though they are by default normalized in the SSD models anyway; meaning the default was double normalization and triple normalization is possible. This change means that coordinates are only normalized if the applyRatio flag is set, which seems appropriate based on the documentation.