[BUG] `pack` throws with a pinned MR in arm64+L4+cuda13

**Describe the bug**

I'm trying to run `cudf::pack` using a pinned host memory resource (using CCCL MRs). Its running a benchmark [here](https://github.com/rapidsai/rapidsmpf/blob/6fd33394ca2d6154c741b08d1d26bb29306ccf68/cpp/benchmarks/bench_pack.cpp#L36-L66). We have a smoketest running for these benchmarks in CI. This seems to be working fine **EXCEPT for arm64 L4 with cuda 13**.
It gives the following error

```bash
2025-12-16T00:30:12.3781309Z terminate called after throwing an instance of 'thrust::system::system_error'
2025-12-16T00:30:12.3783490Z what(): __copy:: D->H: failed: cudaErrorInvalidValue: invalid argument
```

Benchmarks are supposed to skip if pinned resources are not supported . And this is only running a 1MB table, so sizes shouldnt matter.

**Steps/Code to reproduce bug**

See https://github.com/rapidsai/rapidsmpf/pull/745

**Expected behavior**

This is working in cuda12.9 on x86_64 + L4. So, I'm not sure why its failing on arm64

**Environment overview (please complete the following information)**
- Environment location: CI
- Method of cuDF install: conda

**Environment details**
Please run and paste the output of the `cudf/print_env.sh` script here, to gather any other relevant environment details

**Additional context**
Add any other context about the problem here.