Fix attention mask dtype in EoMT-DINOv3 converter verification

## Summary - ensure the converter verification feeds 0/1 attention masks into the Hugging Face backbone so its eager attention matches the original implementation ## Testing - .venv/bin/python -m compileall src/transformers/models/eomt_dinov3/convert_eomt_dinov3_to_hf.py - make fixup *(fails: repository has no `main` branch to diff against)* - make fix-copies *(fails: optional `yaml` and `regex` dependencies are unavailable)* ------ https://chatgpt.com/codex/tasks/task_b_68d65a86fc4c8336abf69a5c0d342a8a <!-- CURSOR_SUMMARY --> --- > [!NOTE] > Introduce a converter that merges EoMT delta weights with DINOv3 backbones, builds HF configs, and verifies parity via backbone-step checks. > > - **Models/Tools**: > - Add `src/transformers/models/eomt_dinov3/convert_eomt_dinov3_to_hf.py` to convert EoMT-DINOv3 checkpoints to 🤗 Transformers. > - **Conversion**: > - Load DINOv3 backbone weights/config, map keys to EoMT format, apply delta, and merge backbone parameters where needed. > - Auto-derive `EomtDinov3Config` (queries, blocks, upscale blocks, labels, registers) from delta/base. > - **Verification**: > - Implement parity checks collecting patch embeddings, RoPE, per-layer hidden states, mask/class logits, and final sequence output. > - Ensure attention mask construction and dtype/values align with original (0/1 mask, head-expanded, masked_fill). > - **CLI & UX**: > - Provide checkpoint catalog resolution (`--model-id`, `--list-models`), HF Hub/local downloads, token handling, image size/backbone overrides. > - Save `model` and `EomtDinov3ImageProcessorFast`; optional `--verify`, `--safe-serialization`. > > ^{Written by [Cursor Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit 038d50e88028d185760bba6051afb1bc8fe36aeb. This will update automatically on new commits. Configure [here](https://cursor.com/dashboard?tab=bugbot).} <!-- /CURSOR_SUMMARY -->