[linux-arm64] Reduce parallelism to 32 for nogil pytest cpu to avoid OOMs.
linux-arm64 nogil pytest cpu fail for both py3.13 and py3.14. Bazel tests, x86 pass.
Example: https://github.com/jax-ml/jax/actions/runs/22460361924/job/65052695600
Nightly presubmit passes for the failing jobs: https://github.com/jax-ml/jax/actions/runs/22463951941
PiperOrigin-RevId: 875940807