ci: balance heavy tests across test matrix legs

[jezdez]

Description

Stacked on top of #5963.

This PR introduces an @pytest.mark.heavy marker and uses it to distribute five high-duration integration tests between the two xdist matrix legs.

The current durations/Linux.json records:

665.9s  test_api_build_grpc_issue5645
185.1s  test_conda_py_no_period
184.9s  test_sysroots_detection
126.0s  test_cross_unix_windows_mingw
124.5s  test_ignore_run_exports_from_substr

These stored call durations total about 1,286 seconds. They overlap when tests run under xdist, so they are useful for selecting expensive tests but are not a wall-time estimate.

Linux and macOS route heavy tests to the matrix leg named serial. Both matrix legs continue to run with pytest -n auto. On Windows, heavy and slow tests remain excluded, preserving the existing exclusion of test_conda_py_no_period.

The intended effect is to distribute long-running tests more evenly between concurrently scheduled CI jobs without changing their execution semantics.

Checklist - did you ...

• Add a file to the news directory (using the template) for the next release's release notes?
• Add / update necessary tests? — adds markers to five existing tests
• Add / update outdated documentation? — no user-facing documentation affected

[codspeed-hq]

Merging this PR will not alter performance

✅ 5 untouched benchmarks

---

_{Comparing ci/heavy-test-marker (baa286f) with ci/test-speed-hygiene (b4f0aa7)}