Reduce EoMT instance validation mask memory

[KrzysiekSko]

What has changed and why?

Implements Task 1 from #941 for the DINOv2 and DINOv3 EoMT instance-segmentation validation paths.

• Processes the query dimension in chunks sized from a 512 MiB estimated working-set budget rather than a fixed query count.
• Accounts conservatively for six FP32-sized mask-score buffers per output pixel.
• Resizes and scores one chunk at a time.
• Moves completed prediction masks to CPU immediately while keeping labels and scores on their original device.
• Leaves get_labels_masks_scores unchanged.
• Adds focused tests for byte-budget chunk sizing, actual chunk boundaries, and numerical equality with the unchunked path.

The metric semantics and original-image evaluation resolution are unchanged.

Refs #941

How has it been tested?

The focused tests are included at:

tests/_task_models/eomt/test_instance_segmentation.py

They compare all labels, masks, and scores from a forced multi-chunk run against a single-chunk run using deterministic tensors.

A full local test run and GPU peak-memory measurement were not available in this environment. CI is requested for the repository test matrix. The device check remaining for review is validation on the same checkpoint/data before and after the change, including torch.cuda.max_memory_allocated() on a wide image.

Did you update CHANGELOG.md?

• Yes
• Not needed (internal change)

Did you update the documentation?

• Yes
• Not needed (internal change without effects for user)

[CLAassistant]

All committers have signed the CLA.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5570d08ac1

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".

[KrzysiekSko]

Addressed in c21701f and formatted in 6015441.

The chunk budget now uses the larger materialized mask surface by pixel count across model_image_size and the final image_size, so the first resize cannot bypass the budget when the final image is smaller. I added a regression where model_image_size=(8, 10) and image_size=(4, 5); a two-query model-resolution budget produces chunk boundaries [2, 2, 1].

Validation:

• uv run pytest tests/_task_models/eomt/test_instance_segmentation.py -q: 3 passed
• focused Ruff check: passed
• focused Ruff format check: passed
• git diff --check: passed

[liopeer]

/review