feat(cli): add rfdetr export subcommand for ONNX/TFLite export

[omkar-334]

What does this PR do?

Adds a command-line path for exporting trained RF-DETR checkpoints, resolving the long-standing TODO in rfdetr/export/main.py to expose TFLite export (and its quantization / calibration options) from the shell. Until now, export was reachable only from Python via RFDETR.export(format=...).

Rather than extend the legacy argparser, this adds a proper rfdetr export subcommand that wraps RFDETR.export (TFLite already worked in Python; this surfaces it to the CLI).

Related to #1024

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:
I verified the training and export end to end -> Train a checkpoint, run rfdetr export --checkpoint <ckpt> --format onnx, and the exported rfdetr-nano.onnx loads and runs in onnxruntime (dets / labels outputs).

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)

Additional Context

export (jsonargparse) and fit/validate/test/predict (LightningCLI) are two different parsers, so a single rfdetr command needs a thin root to choose between them. Treating LightningCLI as the root would force the [train] extra on export-only users and produce a stitched, two-part top-level --help. The root therefore owns top-level help and dispatch and delegates to each backend, which leaves LightningCLI's training semantics untouched and keeps per-command rfdetr <command> --help backend-native.

Usage

# install (no [train] / PyTorch Lightning needed for export)
pip install "rfdetr[cli,onnx]"          # ONNX
pip install "rfdetr[cli,onnx,tflite]"   # TFLite

# discover
rfdetr --help            # lists fit/validate/test/predict + the export subcommand
rfdetr export --help     # all flags, derived from RFDETR.export

# ONNX
rfdetr export --checkpoint output/checkpoint_best_total.pth --format onnx --output_dir output

# TFLite (FP32 + FP16)
rfdetr export --checkpoint output/checkpoint_best_total.pth --format tflite --output_dir output

# TFLite INT8 with calibration images
rfdetr export --checkpoint output/checkpoint_best_total.pth --format tflite --quantization int8 --calibration_data path/to/val_images/ --max_images 100 --output_dir output

# flags from a YAML file
rfdetr export --config export.yaml

[codecov]

Codecov Report

❌ Patch coverage is 92.30769% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 79%. Comparing base (779e552) to head (fc750ae).
⚠️ Report is 34 commits behind head on develop.

❌ Your project check has failed because the head coverage (79%) is below the target coverage (95%). You can increase the head coverage or adjust the target coverage.

Additional details and impacted files

@@           Coverage Diff            @@
##           develop   #1067    +/-   ##
========================================
- Coverage       81%     79%    -2%     
========================================
  Files          108     109     +1     
  Lines        11215   11264    +49     
========================================
- Hits          9120    8906   -214     
- Misses        2095    2358   +263     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Borda]

Lte's do not add it until it is stable, so far I tested TFlite, the predictions were quite questionable...

[Borda]

FYI, we have moved the TensorRT export from subcomand to native package use in #853...

[Copilot]

Pull request overview

This PR adds a unified rfdetr root CLI that lazily dispatches to either the training/eval LightningCLI backend or a new rfdetr export subcommand (jsonargparse) that wraps RFDETR.export, enabling ONNX/TFLite export (including TFLite quantization/calibration flags) from the shell without requiring the [train] extra.

Changes:

• Introduces rfdetr export (src/rfdetr/cli/export.py) and a root dispatcher (src/rfdetr/cli/__init__.py) to route commands without eagerly importing the training stack.
• Moves the LightningCLI backend to src/rfdetr/cli/train.py and updates imports/tests/docs accordingly; adds a lazy re-export for RFDETRCli from rfdetr.training.
• Updates documentation and configuration to reflect the new CLI structure and adds CLI routing/forwarding tests.

Reviewed changes

Copilot reviewed 10 out of 11 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/cli/test_train.py	Updates CLI import-path expectations to rfdetr.cli.train.
tests/cli/test_smoke.py	Updates smoke helper to import RFDETRCli from rfdetr.cli.train.
tests/cli/test_export.py	Adds tests for rfdetr export forwarding and root-dispatch behavior (argv handling, help, invalid command).
src/rfdetr/training/init.py	Lazily re-exports RFDETRCli to avoid circular imports while preserving rfdetr.training.RFDETRCli.
src/rfdetr/export/main.py	Updates comments to point users to the new rfdetr export --format tflite path for TFLite export.
src/rfdetr/cli/train.py	Adds the LightningCLI backend module (fit/validate/test/predict).
src/rfdetr/cli/export.py	Adds rfdetr export jsonargparse wrapper over from_checkpoint(...).export(...).
src/rfdetr/cli/init.py	Implements the root CLI help + lazy dispatch to training backend or standalone commands.
pyproject.toml	Ensures rfdetr console script points to rfdetr.cli:main and updates tooling overrides for moved modules.
docs/reference/training.md	Updates mkdocstrings reference to rfdetr.cli.train.RFDETRCli.
docs/learn/export.md	Documents the new command-line export workflow and examples.