coco mask converting

[IgorDavidyuk]

vision/references/detection/coco_utils.py

Line 170 in 4ec38d4

masks = masks.permute(0, 2, 1).contiguous().permute(0, 2, 1)

Docs for MASK R-CNN claims that target should contain 'masks (UInt8Tensor[N, 1, H, W])'
But the line of code above treats it like tensor with 3 dimensions. Catching errors.

[fmassa]

Good catch! This is indeed a bug in the documentation, and was introduced in #1089

I'm sending a PR fixing this