Proper fix for loess linalg error - non-random tests (Ouranosinc#2020)

### What kind of change does this PR introduce?

* Non-random tests are, in my opinion, better than random tests. This
fixes the data for test_loess_smoothing_nan to be something known to
work. Loess_smoothing does not work on singular matrices (which are very
rare to occur randomly, but *do* occur).
* Also removes the too harsh condition on np.linalg.LinAlgError, which
could occur for other reasons than singular matrices.

tests/test_sdba/test_loess.py

@@ -1,7 +1,6 @@
+# ruff: noqa: E241
 from __future__ import annotations
 
-import logging
-
 import numpy as np
 import pandas as pd
 import pytest
@@ -13,7 +12,7 @@
     _linear_regression,  # noqa
     _loess_nb,  # noqa
     _tricube_weighting,  # noqa
-    loess_smoothing,
+    loess_smoothing,  # noqa
 )
 
 
@@ -74,7 +73,22 @@ def test_loess_smoothing(use_dask, open_dataset):
 @pytest.mark.parametrize("use_dask", [True, False])
 def test_loess_smoothing_nan(use_dask):
     # create data with one axis full of nan
-    data = np.random.randn(2, 2, 10)
+    # (random array taken from np.random.randn)
+    # fmt: off
+    data = np.array(
+        [
+            -0.993, -0.980, -0.452, -0.076,  0.447,
+             0.389,  2.408,  0.966, -0.793,  0.090,
+            -0.173,  1.713, -1.579,  0.454, -0.272,
+            -0.005, -0.926, -2.022, -1.661, -0.493,
+            -0.643,  0.891,  0.194,  0.086,  0.983,
+            -1.048,  2.032,  1.174, -0.441, -0.204,
+            -1.126,  0.933,  1.987,  0.638,  0.789,
+             0.767,  0.676, -1.028,  1.422,  0.453,
+        ]
+    )
+    # fmt: on
+    data = data.reshape(2, 2, 10)  # pylint: disable=too-many-function-args
     data[0, 0] = [np.nan] * 10
     da = xr.DataArray(
         data,
@@ -86,11 +100,4 @@ def test_loess_smoothing_nan(use_dask):
 
     assert out.dims == da.dims
     # check that the output is all nan on the axis with nan in the input
-    try:
-        assert np.isnan(out.values[0, 0]).all()
-    except np.linalg.LinAlgError:
-        msg = (
-            "This has roughly a 1/50,000,000 chance of occurring. Buy a lottery ticket!"
-        )
-        logging.error(msg)
-        pass
+    assert np.isnan(out.values[0, 0]).all()