Add CUTLASS support and mark Evoformer test as sequential
- Install CUTLASS v3.5.1 and set CUTLASS_PATH for Evoformer op
- Mark test_DS4Sci_EvoformerAttention as sequential to avoid CUDA fork issues
- Restore full test workflow with all dependencies
Signed-off-by: Masahiro Tanaka <mtanaka@anyscale.com>