feat(tflite): add private inference helpers and tests - #1011

This pull request introduces a new TFLite inference helper module for RF-DETR exported models, along with a comprehensive test suite. The main focus is to provide lightweight, dependency-minimal utilities for running inference on TFLite models outside the training stack, and to ensure robust and reliable operation through detailed unit tests.

Key additions and improvements:

TFLite Inference Utilities

• Added src/rfdetr/export/_tflite/inference.py, which provides helper functions for TFLite model inference:
  • _create_interpreter: Loads and prepares a TFLite interpreter, supporting both tflite-runtime and tensorflow.lite as backends, and prints input/output tensor shapes.
  • _run_inference: Handles image preprocessing (including normalization and resizing), runs inference, and decodes detection outputs into pixel-space bounding boxes using the supervision library. Supports both RGB and grayscale images and is robust to output tensor ordering.
  • Includes a numerically stable _softmax function for post-processing logits.

Comprehensive Testing

• Added tests/export/test_tflite_inference.py with detailed unit tests for the new inference helpers:
  • Tests for interpreter creation, backend fallback logic, tensor allocation, and correct handling of both string and Path model paths.
  • Tests for inference correctness, including detection thresholding, pixel-space box scaling, handling of RGB and grayscale images, and robustness to output tensor reordering.
  • Uses extensive mocking to simulate TFLite interpreters and image inputs, ensuring tests do not require actual TFLite models or hardware.

These changes make it easy to deploy and test RF-DETR TFLite models in environments where only minimal dependencies are available, and ensure correctness and robustness through thorough testing.