fix(linalg.inv): zero output on failure to prevent silent data corruption

When blas_ops::inv returns false (LAPACK unavailable or singular),
the output buffer previously held the unmodified copy of A — which
callers could mistake for a valid inverse.  Now zeros the output
on failure so the caller sees unambiguous garbage rather than the
original matrix disguised as the inverse.

Callers MUST check the return value.  The pybind11 wrapper already
throws std::runtime_error on failure.

Author: peng.li24

numpycpp/linalg.h

@@ -99,11 +99,17 @@ inline void norm_axis(const T* src, T* dst,
 /// numpy.linalg.inv(a) — matrix inverse (square N×N)
 /// Uses DGESV (bitexact) or Gauss-Jordan (std backend).
 /// Returns true on success; false if matrix is singular or LAPACK unavailable.
+/// On failure, A_inv is zeroed (caller MUST check the return value).
 template<typename T>
 inline bool inv(const T* A, T* A_inv, size_t N) {
-    // Copy input to output buffer (inv modifies in-place)
+    // Copy input to output buffer (blas_ops::inv modifies in-place)
     for (size_t i = 0; i < N * N; ++i) A_inv[i] = A[i];
     bool ok = numpy::detail::blas_ops<T>::inv(A_inv, N);
+    if (!ok) {
+        // Zero output on failure — prevents the caller from reading
+        // the unmodified copy of A as if it were the inverse.
+        for (size_t i = 0; i < N * N; ++i) A_inv[i] = T(0);
+    }
     return ok;
 }