BUG: always return the prediction brush (or mask) of the first bounding box

[yummy527]

Describe the bug
A clear and concise description of what the bug is.
I want to get several masks from an image by adding several bounding boxes.

However, when I trying to do this in Label Studio, I found that: no matter how many boxes were drawn, the prediction returned would always related to the first box.

To Reproduce
Steps to reproduce the behavior:

1. Go to an existing project.
2. Click on any image task. Turn on "Auto-Annotation" and "Auto-Accept Suggestions".
3. Click "Auto-detect" and choose "Rectangle Box". Draw a box, and wait for the prediction. If everything goes well, your will get a predicted mask.
4. Draw the another box, and wait for the prediction. Now, the bug is here: this returned mask corresponds to the first box but not the second box.

Expected behavior
The second predicted mask is expected to correspond to the second box.

Environment (please complete the following information):

• OS: [Ubuntu 22]
• Label Studio Version [1.12.1]

Additional context
If only one box exists, the prediction would always corresponds to that box. That is to say, we could get the right prediction by manually deleting the existed boxes before drawing a new one.

[jombooth]

/jira create

Workflow run
Jira issue TRIAG-760 is created

[jombooth]

Hi @yummy527 - thanks for this report! We'll investigate.