Lite Unify Metric Formatting

[czaloom]

Changes

• IoU changed to IOU for uniformity across valor.
• Reorganized Metric creation into <tasktype>/utilities.py.
• No longer defining a dataclass per metric. Only a single Metric type per task with multiple classmethods that help with initialization.
• Classification and Object Detection now specify maximum_number_of_examples in the ConfusionMatrix metric.
• Classification now follows the same format as object detection and semantic segmentation.
  • A metric value can no longer be a list. A metric value can be either a int, float or dict.
• Remove as_dict parameter to make output typing consistent.

Old Classification Format

{
    "type": "Accuracy",
    "value": [2 / 3, 1 / 3],
    "parameters": {
        "score_thresholds": [0.25, 0.75],
        "hardmax": True,
    },
},

New Classification Format

{
    "type": "Accuracy",
    "value": 2 / 3,
    "parameters": {
        "score_threshold": 0.25,
        "hardmax": True,
    },
},
{
    "type": "Accuracy",
    "value": 1 / 3,
    "parameters": {
        "score_threshold": 0.75,
        "hardmax": True,
    },
},