Run GPU evaluations directly from prepared environment interpreters

Use required project cache for TMPDIR and TRITON_CACHE_DIR in all jobs

Create project envs in CPU setup and separate each baseline smoke evaluation