fix: restore standalone GPU trainer launch

[LeonEricsson]

Summary

background

The README’s environment setup check:

  uv run trainer @ configs/debug/fake/rl.toml

currently fails during distributed initialization due to missing environment variables in os.environ (also noted here. Standalone launches used to work because trainer/envs.py called set_defaults at import time, writing RANK etc into os.environ. #2193 deleted that and inlined os.environ.get("RANK", "0") into World`.

proposal
Rather than reinstate env mutation setup_torch_distributed now passes an in-process HashStore with rank=0, world_size=1 when no launcher is present.

An alternative would be to make the trainer entrypoint a launcher that re-execs under torchrun the way sft and rl already do.