Trim process heap after document analysis

[medisean]

Summary

• Return unused glibc heap pages to the operating system after memory cleanup.
• Run the existing cleanup path after synchronous and asynchronous VLM document analysis.
• Add regression tests for the optional allocator hook and its configuration.

Root cause

Python garbage collection and device-cache cleanup do not necessarily return freed native heap pages to the operating system, so long-running API processes can show increasing RSS across documents.

Validation

• python -m pytest -q tests/unittest/test_model_utils.py — 4 passed
• python -m compileall -q mineru/utils/model_utils.py mineru/backend/vlm/vlm_analyze.py tests/unittest/test_model_utils.py
• Ruff E/F/W checks for the new test file
• git diff --check

Fixes #5313

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[medisean]

I have read the CLA Document and I hereby sign the CLA

[minbaby]

这个 PR 好像没有 Reviewers