BUG: fix nogil LinearNDInterpolator

* Fixes scipygh-21885

* A `nogil` block in `LinearNDInterpolator` is reliably
segfaulting on x86_64 Ubuntu Linux with Python 3.10 and
Python 3.11. I see the same error locally as reported in CI,
and it is present in both `v1.14.1` and on latest `main`
branch of SciPy. Diffing the last succeeding version
of the CI job with the current failing CI job only shows
`tomli` and `hypothesis` changing versions. So, this is pretty
weird, but the test is behind a `slow` marker so maybe we're just
discovering the instability now for some reason.

* Removing the `nogil` directive fixes the issue on Python `3.10`
and `3.11` locally.

* Note that GNU compiler toolchains 9.x and 11.x both reproduce
the segfault locally as well, so the "minimum version" CI label may
be a red herring. It is also worth noting that the CI job name
contains "fast," and the `pytest` incantation isn't using slow,
so there may be yet another issue about i.e., respecting markers
when running `pytest` outside the root folder in that job or something
like that. Nonetheless, the test fails consistently before this fix
locally, which should probably supersede performance concerns for now.

[skip circle] [skip cirrus]

Author: tylerjereddy

scipy/interpolate/_interpnd.pyx

@@ -336,30 +336,32 @@ class LinearNDInterpolator(NDInterpolatorBase):
         eps = 100 * DBL_EPSILON
         eps_broad = sqrt(DBL_EPSILON)
 
-        with nogil:
-            for i in range(xi.shape[0]):
+        # NOTE: a nogil block segfaults here with Python 3.10
+        # and 3.11 on x86_64 Ubuntu Linux with gcc 9.x and 11.x
+        # and therefore nogil was disabled to fix gh-21885
+        for i in range(xi.shape[0]):
 
-                # 1) Find the simplex
+            # 1) Find the simplex
 
-                isimplex = qhull._find_simplex(&info, c,
-                                               &xi[0,0] + i*ndim,
-                                               &start, eps, eps_broad)
+            isimplex = qhull._find_simplex(&info, c,
+                                           &xi[0,0] + i*ndim,
+                                           &start, eps, eps_broad)
 
-                # 2) Linear barycentric interpolation
-
-                if isimplex == -1:
-                    # don't extrapolate
-                    for k in range(nvalues):
-                        out[i,k] = fill_value
-                    continue
+            # 2) Linear barycentric interpolation
 
+            if isimplex == -1:
+                # don't extrapolate
                 for k in range(nvalues):
-                    out[i,k] = 0
+                    out[i,k] = fill_value
+                continue
 
-                for j in range(ndim+1):
-                    for k in range(nvalues):
-                        m = simplices[isimplex,j]
-                        out[i,k] = out[i,k] + c[j] * values[m,k]
+            for k in range(nvalues):
+                out[i,k] = 0
+
+            for j in range(ndim+1):
+                for k in range(nvalues):
+                    m = simplices[isimplex,j]
+                    out[i,k] = out[i,k] + c[j] * values[m,k]
 
         return out