Implement loop peeling from back

As discussed today IRL. A copy of peel_loop_front mutatis mutandis.

src/pystencils/backend/transformations/reshape_loops.py

@@ -48,7 +48,9 @@ class ReshapeLoops:
             peeled_ctr = self._factory.parse_index(
                 cc.get_replacement(loop.counter.symbol)
             )
-            peeled_idx = self._typify(loop.start + PsExpression.make(PsConstant(i)))
+            peeled_idx = self._elim_constants(
+                self._typify(loop.start + PsExpression.make(PsConstant(i)) * loop.step)
+            )
 
             counter_decl = PsDeclaration(peeled_ctr, peeled_idx)
             peeled_block = self._canon_clone.visit(loop.body, cc)
@@ -65,7 +67,9 @@ class ReshapeLoops:
             peeled_iters.append(peeled_block)
 
         loop.start = self._elim_constants(
-            self._typify(loop.start + PsExpression.make(PsConstant(num_iterations)))
+            self._typify(
+                loop.start + PsExpression.make(PsConstant(num_iterations)) * loop.step
+            )
         )
 
         return peeled_iters, loop
@@ -88,6 +92,13 @@ class ReshapeLoops:
             Tuple containing the modified loop and the peeled-off iterations (sequence of blocks).
         """
 
+        if not (
+            isinstance(loop.step, PsConstantExpr) and loop.step.constant.value == 1
+        ):
+            raise NotImplementedError(
+                "Peeling iterations from the back of loops is only implemented for loops with unit step. Implementation is deferred until loop range canonicalization is available (also needed for the vectorizer)."
+            )
+
         peeled_iters: list[PsBlock] = []
 
         for i in range(num_iterations)[::-1]: