Fix inconsistent performance when infer SVTR ONNX

[thsno02]

Re-submit the PR: [Fix inconsistent performance when infer SVTR ONNX](#9371 (comment))

Fix inconsistent performance when inferring SVTR ONNX:

• Paddle OCR 的训练模型转换为ONNX模型后识别效果发生改变，原来可以正确识别的，转换后识别错误了，请问是什么原因？ #8149
• 自己训练模型。有这样一个有趣现象。 不知道怎么解决。 转换成onnx模型，用于截屏识别 文字放大时截图的效果有时比原始图好，不知道怎样自动放大，确保识别效果最好 微信nlanguage #7551

[paddle-bot]

Thanks for your contribution!

[Copilot]

Pull Request Overview

This PR addresses inconsistent performance when inferring SVTR ONNX by modifying the image resize normalization function.

• Updates the resize_norm_img_svtr function to adjust the image width dynamically based on the input aspect ratio.
• Introduces padding for the resized image to ensure consistent dimensions.

Comments suppressed due to low confidence (1)

tools/infer/predict_rec.py:387

• [nitpick] While not necessarily a naming error, the conditional logic determining resized_w could benefit from inline comments or a more descriptive variable name to clarify its purpose.

if math.ceil(imgH * ratio) > imgW:

[Copilot]

The computation for max_wh_ratio is redundant since min(max(max_wh_ratio, ratio), max_wh_ratio) will always evaluate to the original max_wh_ratio. Consider removing this line or revising it to correctly adjust the ratio if that was the intent.

Suggested change

	max_wh_ratio = min(max(max_wh_ratio, ratio), max_wh_ratio)
	max_wh_ratio = max(max_wh_ratio, ratio)