ci: run the windows python tests on two xdist workers

[lstein]

Summary

The Windows python-test jobs flake non-deterministically. They ran pytest -n logical — one xdist worker per vCPU, four on a windows-2022 standard runner. OMP_NUM_THREADS/MKL_NUM_THREADS cap torch's pools at one thread per worker, but they do not cap the helper processes some tests spawn (ffmpeg decoders in tests/app/util/test_video_thumbnails.py, installer subprocesses), so a full-width run leaves the runner with no spare core and the deadline-sensitive tests miss their budgets at random.

Observed on recent runs:

• tests/app/util/test_video_thumbnails.py::TestStreamedDecoderIsBounded::test_capacity_wait_and_first_frame_share_one_deadline — first frame got a fresh timeout after the capacity wait (2.89s / 3.58s)
• tests/app/services/shared/test_for_scheduler_performance.py::test_loop_scheduler_overhead_is_linear[iterate]

Windows now runs two workers. The worker count and the job timeout moved into the matrix; linux and macos keep -n logical and the 30-minute timeout. Windows gets 45 minutes, because the pytest step was already ~13.5 min of a ~15 min job at four workers and halving the workers does not leave room under 30.

Not fixed here, and still open: the second Windows failure mode on these runs is a hard Windows fatal exception: code 0xc000001d (illegal instruction) inside F.linear on bfloat16 CPU tensors — invokeai/backend/quantization/dequantizing_linear.py:49, reached from the NVFP4 loader tests (test_krea2_nvfp4_loader.py, test_z_image_loader_boundaries.py). It is a CPU-ISA dispatch crash, not contention, and it is non-deterministic because GitHub's Windows pool mixes CPU generations. Fewer workers does not address it.

QA Instructions

• python -c "import yaml; yaml.safe_load(open('.github/workflows/python-tests.yml'))" — parses.
• Effect is observable only on the hosted Windows runners; this PR's own windows-cpu jobs are the check. Confirm both Windows jobs run -n 2 and finish inside 45 minutes.
• Baseline measured from run 35675232701 (py3.11: windows-cpu, passing): run pytest 809s, whole job ~14 min.

Review

Mechanical CI configuration change; no independent review subagents were run. Self-checks: YAML parses and the matrix resolves to the intended per-platform values; release.yml calls this workflow via workflow_call and does not touch the matrix, so it is unaffected.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

🤖 Generated with Claude Code