test(cfo): use i32 offsets in dynamic block pointers

Keep descriptor storage wide while satisfying Triton make_block_ptr's 32-bit offset contract in both if branches. This fixes the Group A pytest input without changing product lowering or marker semantics.

Author: elstainniles

third_party/ascend/unittest/pytest_ut/test_control_flow_pointer_boundary.py

@@ -35,8 +35,11 @@ def block_ptr_if_descriptor_kernel(a, b, out, choose_ptr, descriptor_ptr, BLOCK:
     size_b = tl.load(descriptor_ptr + 1)
     stride_a = tl.load(descriptor_ptr + 2)
     stride_b = tl.load(descriptor_ptr + 3)
-    offset_a = tl.load(descriptor_ptr + 4)
-    offset_b = tl.load(descriptor_ptr + 5)
+    # Block-pointer offsets are required to use Triton's 32-bit index type.
+    # Keep the descriptor storage wide, but narrow the values at the frontend
+    # boundary instead of changing the product's make_block_ptr contract.
+    offset_a = tl.load(descriptor_ptr + 4).to(tl.int32)
+    offset_b = tl.load(descriptor_ptr + 5).to(tl.int32)
     if choose != 0:
         pointer = tl.make_block_ptr(
             base=a + 3,