Refactor Req.fill_ids into full_untruncated_fill_ids + fill_len with equivalence (#26637)

test/registered/unit/utils/test_common.py

@@ -14,8 +14,8 @@ register_cuda_ci(est_time=5, stage="base-b", runner_config="1-gpu-small")
 class TestFlattenArraysToInt64Tensor(CustomTestCase):
     """`flatten_arrays_to_int64_tensor` is invoked by `prepare_for_extend`
     to build the per-batch input_ids tensor (pinned, async H2D) from a
-    list of array.array('q') per-req fill_ids slices. Tests the full
-    matrix of (device, pin) the production code paths through.
+    list of array.array('q') per-req get_fill_ids() slices. Tests the
+    full matrix of (device, pin) the production code paths through.
     """
 
     DEVICES = ("cpu", "cuda")