fix(seg): filter unlabeled images and fix mask-loss DDP graph disconnect

[imagra93]

Problem

DDP segmentation training (static_graph=True) crashes when the dataset contains unlabeled images. Two independent code paths contribute:

1. Disconnected gradient in loss_masks — when a rank's batch has
   zero Hungarian matches, the factored mask-head branch returns
   torch.tensor([], device=device). This tensor has no grad_fn, so
   spatial_features, query_features, and bias are silently dropped
   from the backward graph on that rank, violating static_graph=True.

2. Unlabeled images in the dataloader — images with no annotation
   file inflate the dataset and can produce all-unlabeled batches on some
   ranks, making the no-match case structurally inevitable.

Fixes #291

Solution

libreyolo/models/rfdetr/loss.py
Return zero losses routed through the mask-head tensors instead of a
disconnected empty tensor:

zero = spatial_features.sum() * 0.0 + query_features.sum() * 0.0 + bias * 0.0
return {"loss_mask_ce": zero, "loss_mask_dice": zero}

Every mask-head parameter participates in backward (with zero gradient),
satisfying DDP's reducer regardless of batch content.

libreyolo/data/dataset.py

New filter_empty_annotations: bool = False parameter on YOLODataset.
When True, zero-annotation images are dropped at construction time. A
warning is logged (main process only) with the count of dropped images.
Raises ValueError if the entire dataset is unlabeled.

libreyolo/training/trainer.py

filter_empty_annotations=load_segments passed to all three
YOLODataset construction sites, enabling the filter automatically for
every segmentation training path.

[imagra93]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Breezy!

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[EHxuban11]

I'm concerned that this PR is dropping the unlabelled images basically (often called background images)d, but background images are useful in order to avoid the model to detect false positives when there isn't an object in the image.