feat(export): add TFLite export support with FP32/FP16/INT8

[mfazrinizar]

What does this PR do?

Adds an ONNX → TFLite conversion pipeline for RF-DETR detection and segmentation models via onnx2tf.

Users can now export any RF-DETR model to TFLite format with a single call:

from rfdetr import RFDETRNano

model = RFDETRNano()

# FP32 (recommended — also produces FP16 automatically)
model.export(format="tflite")

# INT8 with calibration images
model.export(format="tflite", quantization="int8", calibration_data="path/to/images/")

Key changes

New module: src/rfdetr/export/_tflite/

• converter.py: Core export_tflite() function — converts ONNX to TFLite via onnx2tf Python API
• Supports quantization="int8" with calibration from an image directory, .npy file, or NumPy array
• max_images parameter controls calibration sample count (default: 100)
• Applies output_signaturedefs=True fix for segmentation models (onnx2tf node naming issue with leading / characters violating TF saved_model naming pattern)
• Monkey-patches onnx2tf's download_test_image_data() to use local calibration data instead of downloading from GitHub (which fails in CI, firewalls, air-gapped systems)
• NumPy allow_pickle compatibility shim for onnx2tf 1.x releases

Modified files

• src/rfdetr/detr.py: Added format="tflite" routing in export() with quantization, calibration_data, max_images, output_dir parameters
• src/rfdetr/export/main.py / __init__.py: Wired up export_tflite import
• pyproject.toml: Added onnx2tf>=1.26.0 and tf-keras>=2.16.0 to [onnx] optional dependencies
• docs/learn/export.md: TFLite usage guide, calibration explanation, parameter reference, FP32/FP16 always-produced behavior note

Related Issue(s): Related to #173 and this PR

Type of Change

• New feature (non-breaking change that adds functionality)

Testing

• I have tested this change locally
• I have added/updated tests for this change

Test details:

56 unit tests in tests/export/test_tflite_export.py (all CPU, no GPU required):

• Format routing: export(format="tflite") correctly dispatches to export_tflite
• Quantization options: Validates None, "fp32", "fp16", "int8" modes; rejects invalid values
• Calibration data preparation: Tests random noise generation, .npy file loading, NumPy array input, and directory-based image loading with resize
• max_images parameter forwarding: Verifies max_images flows through detr.py → export_tflite() → _prepare_calibration_data() → _load_calibration_images()
• Error handling: Missing ONNX file, missing calibration directory, empty image directory
• INT8 warning: Emits warning when no calibration data provided for INT8

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• I have updated the documentation accordingly (if applicable)

[CLAassistant]

All committers have signed the CLA.