Performance regression in np.random.RandomState() skews performance test results across Python 2/3 on Dataflow

Tests show that the performance of core Beam operations in Python 3.x on Dataflow can be a few time slower than in Python 2.7. We should investigate what's the cause of the problem.

Currently, we have one ParDo test that is run both in Py3 and Py2 [1]. A dashboard with runtime results can be found here [2].

[1] sdks/python/apache_beam/testing/load_tests/pardo_test.py

[2] https://apache-beam-testing.appspot.com/explore?dashboard=5678187241537536