BUG: failing `sqrtm` regression test

[rgommers]

Caused by gh-20212. Marked as xfail in gh-20218. Failing with both OpenBLAS 0.3.21 and new Accelerate - see #19816 (comment).

From this CI log:

____________________________ TestSqrtM.test_gh17918 ____________________________
  ../venv-test/lib/python3.12/site-packages/scipy/linalg/tests/test_matfuncs.py:423: in test_gh17918
      assert np.isrealobj(sqrtm(M))
  E   assert False
  E    +  where False = <function isrealobj at 0x114d80e30>(array([[0.45485703-1.38777878e-17j, 0.20990806+3.46944695e-17j,\n        0.20990806-1.08420217e-18j, 0.20990806-1.30104...+1.08420217e-19j,\n        0.20990806+1.08420217e-19j, 0.20990806+1.08420217e-19j,\n        0.45485703+1.08420217e-19j]]))
  E    +    where <function isrealobj at 0x114d80e30> = np.isrealobj
  E    +    and   array([[0.45485703-1.38777878e-17j, 0.20990806+3.46944695e-17j,\n        0.20990806-1.08420217e-18j, 0.20990806-1.30104...+1.08420217e-19j,\n        0.20990806+1.08420217e-19j, 0.20990806+1.08420217e-19j,\n        0.45485703+1.08420217e-19j]]) = sqrtm(array([[1.  , 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94,\n        0.94, 0.94, 0.94, 0.94, 0.94, 0.94, ..., 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94,\n        0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 1.  ]]))
          M          = array([[1.  , 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94,
          0.94, 0.94, 0.94, 0.94, 0.94, 0.94, ..., 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94,
          0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 0.94, 1.  ]])
          self       = <scipy.linalg.tests.test_matfuncs.TestSqrtM object at 0x12b4b79b0>

FAILED linalg/tests/test_matfuncs.py::TestSqrtM::test_gh17918 - assert False

Cc @ev-br