Override linop instead of _evaluate_linop

In arithmetic fallback covariance functions, override `linop` instead of
`_evaluate_linop`, because the latter has the problem that the input
preprocessing destroys attributes of np.array subclasses. Keeping the
option of np.array subclassing can be very useful for certain linop
implementations, e.g. tensor product covariance functions.

Author: timweiland

src/probnum/randprocs/covfuncs/_arithmetic_fallbacks.py

@@ -55,8 +55,8 @@ def __init__(self, covfunc: CovarianceFunction, scalar: ScalarLike):
     def _evaluate(self, x0: np.ndarray, x1: Optional[np.ndarray] = None) -> np.ndarray:
         return self._scalar * self._covfunc(x0, x1)
 
-    def _evaluate_linop(
-        self, x0: np.ndarray, x1: Optional[np.ndarray]
+    def linop(
+        self, x0: utils.ArrayLike, x1: Optional[utils.ArrayLike] = None
     ) -> linops.LinearOperator:
         return self._scalar * self._covfunc.linop(x0, x1)
 
@@ -82,7 +82,6 @@ class SumCovarianceFunction(CovarianceFunction):
     """
 
     def __init__(self, *summands: CovarianceFunction):
-
         if not all(
             (summand.input_shape == summands[0].input_shape)
             and (summand.output_shape_0 == summands[0].output_shape_0)
@@ -104,8 +103,8 @@ def _evaluate(self, x0: np.ndarray, x1: Optional[np.ndarray]) -> np.ndarray:
             operator.add, (summand(x0, x1) for summand in self._summands)
         )
 
-    def _evaluate_linop(
-        self, x0: np.ndarray, x1: Optional[np.ndarray]
+    def linop(
+        self, x0: utils.ArrayLike, x1: Optional[utils.ArrayLike] = None
     ) -> linops.LinearOperator:
         return functools.reduce(
             operator.add, (summand.linop(x0, x1) for summand in self._summands)
@@ -151,7 +150,6 @@ class ProductCovarianceFunction(CovarianceFunction):
     """
 
     def __init__(self, *factors: CovarianceFunction):
-
         if not all(
             (factor.input_shape == factors[0].input_shape)
             and (factor.output_shape_0 == factors[0].output_shape_0)