test: Rebalance parallel test scheduling and trim fixture overhead

[vdusek]

The unit suite spends most of its wall-clock waiting. On a reference CI job (ubuntu-latest / 3.14, run 33047203526) the main pass does 578 worker-seconds of work in 131.8s across 8 workers and burns 476 of them idle. --dist load seeds each worker with a fixed chunk of consecutive tests and never reassigns, so gw4 was still running _playwright/test_utils.py at 125s with the other seven idle for about a minute. Perfect balance would finish at 72.3s.

Same failure mode as apify/apify-client-python#1037.

Changes

1. --dist worksteal in addopts so an idle worker takes over tests still queued on a busy one. Nothing in the repo uses xdist_group, and --dist is inert without -n.
2. proxy and disabled_proxy are session-scoped and sync. Each teardown blocks ~1s on proxy.py's hardcoded selector.select(timeout=1); with 17 servers built and destroyed that is ~17 worker-seconds of pure waiting. The server is a stateless forwarder, so one per session is enough. Sync because a session-scoped async fixture is incompatible with asyncio_default_fixture_loop_scope = "function", and neither ever awaited.
3. Dropped --numprocesses=1 from the run_alone pass - it paid the full xdist worker bootstrap for a 15-test serial suite.
4. COVERAGE_CORE=sysmon on unit-tests-cov. Measured on a 1397-test subset: default core +14%, sysmon +0%. Coverage falls back with a warning where sys.monitoring is unavailable, so 3.10 and 3.11 are safe.
5. test_infinite_scroll_on_dynamic_page waits on the page instead of the clock - four wait_for_timeout(1000) sleeps for a loadMore() that awaits 100ms, on the critical-path worker.

Measured

Full suite, both passes, same machine, two reps each: ~114s -> ~74s (-35%). These runs are without coverage, so change 4 is not reflected. 2456 tests pass on both sides; ruff format, ruff check and ty are clean.

Not in this PR

• from crawlee.crawlers import HttpCrawler costs 3.51s, of which 1.28s is scikit-learn and scipy, imported eagerly by _adaptive_playwright._rendering_type_predictor (behind the optional adaptive-crawler extra). Every worker pays it, as does every crawlee[all] cold start. Not a one-liner - RenderingTypePredictorState needs LogisticRegression at class-creation time. Worth its own issue.
• 14 of the 15 matrix jobs compute coverage that is never uploaded; skipping it needs a change in apify/workflows.
• tests_concurrency: "8" may be 2x oversubscription on 4-vCPU runners; worth one run with nproc echoed and an -n4 comparison.
• test_http_2 still hits the real internet - it needs an HTTP/2 endpoint the local uvicorn server does not speak.

✍️ Drafted by Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.73%. Comparing base (24a4bc5) to head (79be4fd).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2199      +/-   ##
==========================================
- Coverage   93.73%   93.73%   -0.01%     
==========================================
  Files         181      181              
  Lines       12825    12825              
==========================================
- Hits        12022    12021       -1     
- Misses        803      804       +1     

Flag	Coverage Δ
unit	93.73% <ø> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.