[TRAIN-7] RuntimeError: No CUDA GPUs are available on A100; fix requires TORCH_CUDA_ARCH_LIST

## Problem

Running `examples/run_grpo.py` on an A100 node produces `RuntimeError: No CUDA GPUs are available`, even though GPUs are accessible and CUDA is in `PATH`. The fix is to export `TORCH_CUDA_ARCH_LIST="8.0"`, but this is not documented anywhere.

## Steps to reproduce

Run `examples/run_grpo.py` on an A100 node without setting `TORCH_CUDA_ARCH_LIST`. Observe the error.

## Suggested fix

Document that `TORCH_CUDA_ARCH_LIST` must be set to match the target GPU architecture. Consider detecting the GPU architecture automatically or setting a sensible default.