Add Per-Class Confidence Filter workflow block

[dcaroboflow]

Summary

Adds a new transformation block roboflow_core/per_class_confidence_filter@v1 that filters detection predictions by applying a different confidence threshold per class.

• Keeps a detection only if its confidence is at least the threshold configured for its class
• Classes not listed in class_thresholds fall back to a configurable default_threshold (defaults to 0.3)
• Supports object detection, instance segmentation, and keypoint detection prediction kinds
• Registered in core_steps/loader.py; appears as Per-Class Confidence Filter under the flow_control section in the Workflows UI

Why a new block

The existing detections_filter block applies a single confidence threshold across all classes — you can filter to a class list or set a global confidence cutoff, but you cannot say "keep person only above 0.98 and car only above 0.5" without writing extremely verbose nested StatementGroups by hand (which the UI's Configure popup does not surface). This block fills that gap with a purpose-built {class_name: threshold} input for the common per-class case.

API

```json
{
"type": "roboflow_core/per_class_confidence_filter@v1",
"name": "filter",
"predictions": "$steps.detector.predictions",
"class_thresholds": {"person": 0.98, "car": 0.5},
"default_threshold": 0.3
}
```

class_thresholds also accepts a workflow input or step output selector pointing to a dictionary.

Tests

• Unit tests (tests/workflows/unit_tests/core_steps/transformations/test_per_class_confidence_filter.py): manifest validation, per-class threshold drops, inclusive boundary, unknown-class fallback, empty predictions, empty thresholds, batch processing
• Integration test (tests/workflows/integration_tests/execution/test_workflow_with_per_class_confidence_filter.py): full ExecutionEngine path with synthetic detections, including default-threshold fallback and multi-image batches

All 15 tests pass locally. `make style` applied.

Test plan

• `pytest tests/workflows/unit_tests/core_steps/transformations/test_per_class_confidence_filter.py` (12 cases)
• `pytest tests/workflows/integration_tests/execution/test_workflow_with_per_class_confidence_filter.py` (3 cases)
• Manually verified end-to-end via local inference server with mounted volume against YOLO-World predictions on a real image; filter drops/keeps detections per-class as configured
• Verified block registers in /workflows/blocks/describe with correct UI metadata (name, section, icon)

[CLAassistant]

All committers have signed the CLA.