[Dataset] Add downloadable OmniMat data URLs

[yansumm]

Summary

• populate the published TSV URLs for OmniMat_QA and OmniMat_CAL
• fall back to prepare_tsv() when no local OmniMat TSV is available
• preserve the existing local-file precedence and ID normalization
• add focused regression tests for the URL contract and download fallback

This PR intentionally does not change the OmniMat prompts or scoring logic. Those concerns are deferred to a separate follow-up.

Dataset scope and publication status

The linked Hugging Face repository currently contains the OmniMat1K release:

• 498 QA examples
• 502 CAL examples
• 1,000 examples in total

This is not the full 3,171-example dataset described in the paper. The Hugging Face repository is currently public and ungated, but its API metadata does not yet declare a license or dataset owner. This PR is therefore opened as a draft pending confirmation from the maintainers/data owner about the intended upstream naming and publication metadata.

Addresses the OmniMat portion of #1614.

Verification

• pytest -q tests/test_omnimat_data_url.py — 2 passed
• fresh temporary LMUData download — 498 QA + 502 CAL loaded successfully
• both published URLs returned HTTP 206 and downloaded successfully
• uvx pre-commit run --files vlmeval/dataset/omnimat.py tests/test_omnimat_data_url.py — passed
• uvx pre-commit run --all-files — blocked by two pre-existing E226 violations on upstream main:
  • vlmeval/dataset/utils/memlens_utils.py:848
  • vlmeval/dataset/utils/mmlongbench_metrics.py:109