Force NCHW tensor order.

[kkeroo]

This PR forces NCHW tensor output when calling getTensor(...) method instead of manually checking the shapes and transposing them inside parsers.

Affected Parsers:

• MLSD parser
• FastSAM parser
• YOLOExtended parser
• SuperAnimal parser: removed unnecessary check for shapes.

One yolo util function (reshape_seg_outputs) is removed - not needed anymore.

Additionally, ImgDetectionExtended's masks parameter now expects int16 and not int8 (introduced in some previous PR) + detection creator function converts masks to int16 instead of int8.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 28.57143% with 5 lines in your changes missing coverage. Please review.

Project coverage is 32.96%. Comparing base (37a2d92) to head (02db019).
Report is 7 commits behind head on main.

✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
depthai_nodes/ml/messages/creators/detection.py	0.00%	2 Missing ⚠️
depthai_nodes/ml/messages/img_detections.py	33.33%	2 Missing ⚠️
depthai_nodes/ml/parsers/mlsd.py	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #135      +/-   ##
==========================================
- Coverage   33.40%   32.96%   -0.45%     
==========================================
  Files          68       68              
  Lines        3739     3695      -44     
==========================================
- Hits         1249     1218      -31     
+ Misses       2490     2477      -13     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jkbmrz]

Looks good but why are we not adding this feature to all parsers?

[kkeroo]

Other parsers dont explicitly check for the tensor order - it doesnt matter. E.g.:

• Classification parser: output is not NCHW and it doesnt make sense to be
• Keypoints parser: automatically reshapes the output based on num_kpts

and similar for other not-affected parsers. I think forcing storageOrder=NCHW in every getTensor method is not needed and not ideal because DAI could fail transforming it if the shapes are weird