Integrate RF-DETR into VisDet

[GeorgePearse]

Summary

Add support for RF-DETR (Roboflow DETR variant) within VisDet so it can be trained and run via the standard APIs (configs/presets + inferencers).

Goals

• Make RF-DETR available as a first-class model option in VisDet
• Support inference via DetInferencer (preset + weights)
• Support training/evaluation via existing runner tooling

Scope / Tasks

• Identify upstream RF-DETR reference implementation and license compatibility
• Add model components (backbone/neck/transformer/head) or adapter layer
• Add config(s) / YAML preset(s) and metadata (classes, img size, etc.)
• Add checkpoint download story (where weights live; HF/GitHub/etc.)
• Add smoke tests (import + forward + minimal infer)
• Document usage (install requirements, inference example, training command)

Acceptance Criteria

• DetInferencer(model="rf-detr-...") runs end-to-end on a sample image
• Training can be launched with a config/preset without manual code changes
• CI has at least a minimal test covering model construction

Notes

• Prefer keeping the integration consistent with VisDet’s registry + preset system
• If extra deps are needed, put them behind an optional extra group