Run AutoSP compile tests sequentially (#8020)
This PR fixes AutoSP graph-capture tests that can fail with [PyTorch's
CUDA bad-fork
guard](https://github.com/deepspeedai/DeepSpeed/actions/runs/26216578448/job/77140396424)
when CUDA is visible under pytest --forked.
Signed-off-by: Masahiro Tanaka <mtanaka@anyscale.com>
