refactor GFL head

[RangiLyu]

Refactor GFL head

1. Using center points to generate grid cells instead of using anchor generator.
2. Directly parse the target assign results with a sample function instead of using Sampler class.
3. Rename some parameters and add more comments for easier understanding.

Notice: The center point generated by the refactored code has an offset of 0.5 in the input coordinates from the original GFLHead implemented in MMDetection.

For example: On the feature map which stride=8, the center point coordinate of the first pixel(grid cell) is (4, 4) in the refactored code. But in the MMDetection implemention, the first center point is (3.5, 3.5). In my opinion, (4, 4) should be the correct coordinate.

This may leads to a slightly mAP drop. On nanodet-m the drop is 0.0017. Just several steps of fine-tuning will fix this.

[RangiLyu]

This problem is caused by the different selection of the origin of the coordinate system. Whether to choose the top left corner as the origin of the coordinate system or the center point of the first pixel. I choose to select the top left corner as the origin point, cause after mapping the pixel center point of the input size back to the original image size, the origin is 0.5 times the distance of the downsampling ratio from the real image edge.