[CI] Flaky test in CI windows-gpu

[Kh4L]

Description

windows-gpu CI has some flaky tests.

03:18:11  FAILED tests/python/gpu/test_operator_gpu.py::test_sparse_nd_pickle - mxnet.b...
03:18:11  FAILED tests/python/gpu/test_operator_gpu.py::test_sparse_nd_save_load[save]
03:18:11  FAILED tests/python/gpu/test_operator_gpu.py::test_sparse_ndarray_load_csr_npz_scipy[save]

Occurrences

https://jenkins.mxnet-ci.amazon-ml.com/job/mxnet-validation/job/windows-gpu/job/master/2524/execution/node/155/log/

[TristonC]

@leezu Could you find someone to help?

[DickJC123]

I ran into failures on these exact 3 tests. I've tracked down the problem: the unittest test_zero_sized_dim does not clean up after itself, and leaves is_np_shape_thread_local_ = 1. Subsequent tests run by the same pytest worker might fail as a result. So I get a failure on:

pytest --verbose -s tests/python/gpu/test_operator_gpu.py::{test_zero_sized_dim,test_sparse_nd_pickle}

but not on:

pytest --verbose -s tests/python/gpu/test_operator_gpu.py::test_sparse_nd_pickle

Looking into the most appropriate fix...