Harden local runner for Windows

[krystophny]

Summary

• use local + zeromq as Windows run defaults instead of fork + memmap
• fall back from os.sched_getaffinity to os.cpu_count when CPU affinity is unavailable
• make forked workers spawn/forkserver-safe by using a top-level worker target and loading includes in the child
• fall back to a Python worker command when the profit-worker console script is not installed
• copy template symlink targets when symlink copying is not supported
• document the native Windows run configuration

Fixes #140

Verification

Test fails on main

$ python -m pytest tests/unit_tests/run/test_runner.py::test_runner -q
_pickle.PicklingError: Can't pickle local object ... ForkRunner.spawn.<locals>.work

Test passes after fix

$ python -m pytest tests/unit_tests/config/test_config.py::test_platform_run_defaults tests/unit_tests/run/test_runner.py::test_available_cpus_without_sched_getaffinity tests/unit_tests/run/test_runner.py::test_local_runner_falls_back_to_python_worker tests/unit_tests/run/test_runner.py::test_runner tests/unit_tests/run/test_command.py::test_template_copies_symlink_targets_on_windows -q
6 passed, 1 skipped, 1 warning in 1.48s