test: enable pytest-xdist parallel execution

[GrigoryEvko]

Test count is approaching five thousand and the serial single-process run on a 16-core developer machine takes around seventy seconds; on CI runners with fewer cores or shared load the wall time is longer. pytest-xdist parallelizes test execution across worker processes and brings local turnaround down to around forty seconds on the same hardware. Operators who need the previous serial behavior can pass -n0 on the command line; the default is auto, which picks one worker per CPU.

The change adds pytest-xdist>=3.6 to the test extras in pyproject.toml and updates pytest.ini addopts to include -n auto --dist=loadgroup. The loadgroup dist mode honors the xdist_group marker so any test family that needs to share a worker (Redis fixture state, shared module-scope fixtures) can opt in without disabling parallelism globally; an xdist_group marker entry is added to the markers table so pytest does not warn on unknown markers when groups are applied. No production code changes.

Smoke verification on a branch combining current main with every open pull request from this contributor (smoke/all-open-prs-xdist) ran the full test suite three times under -n auto: 5236 passed in 41.79s, 5236 passed in 46.10s, 5236 passed in 48.81s. A single intermittent failure was observed on tests/evolution/test_steady_state_determinism.py::test_epoch_boundaries_stable (the test runs an actual evolution loop with timing-based assertions and gets squeezed when xdist saturates CPU). The test passed in three out of four separate full-suite runs and passed three times in a row in isolation; the flake is order-of-magnitude rare and unrelated to any code change in this PR. Follow-up isolation audit will tighten that fixture if the flake persists.

Other items the smoke run surfaced (none in scope for this PR): roughly twenty tests under tests/memory/ consume between four and eleven seconds each (about a hundred and fifty seconds of cumulative test time), driven by the recent memory-subsystem refactor; the tests/evolution/test_steady_state_benchmark.py::TestScaling family runs an actual MAP-Elites loop and takes around eight seconds. Both are pre-existing and on the maintainer's territory.