[antonwolfy]

This PR fixes several test infrastructure issues, one commit per fix:

1. -m "slow or not slow" skipped slow tests. pytest_configure already sets -m "not slow" by default, which deselects slow tests. A second check in pytest_collection_modifyitems added a skip mark to slow tests whenever the -m expression contained the substring "not slow", so -m "slow or not slow", documented as the way to run all tests, still skipped them. The redundant check is removed; the default run and -m slow behave as before.
2. Large-allocation tests failed instead of skipping. Third-party tests that allocate huge arrays skip when the allocation fails, but they caught MemoryError. dpnp raises USMAllocationError, which does not derive from MemoryError, so on devices with a small maximum allocation size they failed. They now catch USMAllocationError.
3. Reference to a removed skip list. conftest.py still tried to load skipped_tests_gpu_no_fp64.tbl on GPUs without fp64 support, although the file was deleted once its last entry was gone.
4. "0" enabled test env flags. DPNP_TEST_ALL_INT_TYPES, DPNP_TEST_FLOAT_16, DPNP_TEST_COMPLEX_TYPES, DPNP_TEST_BOOL_TYPES and DPNP_INFRA_WARNINGS_ENABLE were read with bool(os.getenv(...)), so any non-empty value, including 0, turned them on. Unset, empty, 0, false, no and off now mean disabled, and any other value still enables them. SKIP_TENSOR_TESTS uses the same parsing.
5. Removed tests_external. The runner is not used by CI, has not worked since NumPy 2.0 (it imports numpy.core._rational_tests), and its skip lists no longer match the node IDs pytest produces. The dpnp/tests/third_party/numpy_ext wrapper, which only called it, is removed too.

Tested locally on an Intel GPU with a 1 GiB maximum allocation size and on the OpenCL CPU device:

• dpnp/tests/test_fft_gil.py: the default run deselects its 2 slow tests, and both -m slow and -m "slow or not slow" run and pass them (before, -m "slow or not slow" skipped them with "Skipping slow tests").
• The 7 large-allocation tests that failed with USMAllocationError on the GPU now skip.
• The full suite collects the same tests as on master, with and without SKIP_TENSOR_TESTS=1.

TestNansumNanprodHuge still fails on such devices, because dpnp.nansum allocates a temporary larger than the device limit inside the test; that needs a separate change.

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to an issue with a reproducer?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?
• Have you checked performance impact of proposed changes?
• Have you added documentation for your changes, if necessary?
• Have you added your changes to the changelog?

[github-actions]

View rendered docs @ https://intelpython.github.io/dpnp/pull/3103/index.html

[github-actions]

Array API standard conformance tests for dpnp=0.21.0dev15=np2py314h8d9cdd5_22 ran successfully.
Passed: 1376
Failed: 0
Skipped: 6

[coveralls]

Coverage is 82.62% — test-infra-fixes into master. No base build found for master.

[vlad-perevezentsev]

LGTM
Thank you @antonwolfy