改进现有的RetinaNet的量化

## 背景

1.实现fully量化的RetinaNet
2.将量化检测模型和vision/detection进行合并

## 任务描述

1.现有RetinaNet受MegEngine版本限制仅仅量化了backbone部分，后续需要一个fully量化的RetinaNet
2.检测量化部分code有部分冗余，需要进行合并

## 目标

将quant retinanet统一合入 vision/detection

Start by comparing the existing quantized RetinaNet implementation with the vision/detection code, focusing on the current backbone-only quantization and duplicated detection quantization code. Done means RetinaNet is fully quantized and the quantized implementation is unified under vision/detection without the noted redundancy.

Written by the indexing model from the issue text.