Display ConfusionMatrix in StdOut (introduces different meaning of rows and cols to some implementations)

[klemen1999]

Purpose

Until now the confusion matrices were only pushed to the tracker and never displayed in the terminal (stdout). This PR changes this by correctly also showing the confusion matrix under the correct node in which it was used.
Additionally in 8f194d0 it also introduces a breaking change where it unifiys what row and column mean in the matrix for all tasks. Until now the torchmetric and our implementation differed here which can be quite confusing.

Specification

None / not applicable
Multiple matrices under same head (with instance segmentation task)

Multiple matrices under different heads

Implementation when rich_logging: False

Dependencies & Potential Impact

None / not applicable
This PR changes what rows and columns mean in order to unify all to Torchmetrics notation. With improved visualization and notes in the README I deem this a worthy change because the alternative of keeping it different just introduces confusion without any reason.

Deployment Plan

None / not applicable

Testing & Validation

None / not applicable

[dtronmans]

I tested with instance segmentation and it works correctly. For instance segmentation we have two sub-metrics, one for detection and one for segmentation, so maybe here in the artifact keys we can just change it to {sub_name}.json instead of confusion_matrix.json? So that they are logged separately as detection_confusion_matrix.json and segmentation_confusion_matrix.json instead of both mapping to confusion_matrix.json

I verified with this script:

config= "configs/crack_instance_segmentation_constantlr.yaml"

model = LuxonisModel(
    config,
    {
        "tracker.is_mlflow": True,
        "tracker.project_name": "cm_test",
        "tracker.run_name": "cm_test_run",
    },
)

artifact_keys = model.get_mlflow_logging_keys()["artifacts"]
cm_keys = [k for k in artifact_keys if "confusion_matrix" in k]

for key in cm_keys:
    print(key)

With the proposed change I have correctly separate detection and segmentation confusion matrices:

test/metrics/300/PrecisionSegmentBBoxHead/detection_confusion_matrix.json
test/metrics/300/PrecisionSegmentBBoxHead/segmentation_confusion_matrix.json
val/metrics/0/PrecisionSegmentBBoxHead/detection_confusion_matrix.json
val/metrics/0/PrecisionSegmentBBoxHead/segmentation_confusion_matrix.json

Whereas keeping it as it is, both map to the same filename:

val/metrics/0/PrecisionSegmentBBoxHead/confusion_matrix.json
val/metrics/109/PrecisionSegmentBBoxHead/confusion_matrix.json
val/metrics/119/PrecisionSegmentBBoxHead/confusion_matrix.json

[klemen1999]

Nice find, addressed in 578aaff

[dtronmans]

LGTM thanks