Optimize CUDA Breakout rollout and training

[michaelthompsonx-lab]

Summary

• move Breakout physics, reset, reward, terminal, and observation generation onto CUDA
• write rollout observations directly into BF16 storage and consume sampled precision actions without an intermediate action cast
• compact brick state into a four-word bitset and use warp-cooperative observation writes
• capture a complete rollout horizon per buffer and tune Breakout to two buffers
• remove redundant saved-activation copies and accumulate the MinGRU highway gradient through GEMM beta=1
• keep MinGRU backward checkpoints in shared memory, fuse Muon slice normalization, and parallelize PPO advantage moments
• add deterministic CPU/CUDA Breakout parity coverage

Runtime changes are marked with fbr: comments for review. CUDA 13/NCCL build compatibility comments are intentionally untagged.

Verification

• ./build.sh breakout --cuda-env-test && ./test_cuda_env (BF16): PASS, 65 envs x 4096 steps, 1658 completed episodes
• ./build.sh breakout --cuda-env-test --float && ./test_cuda_env (FP32): PASS
• ./build.sh breakout: PASS
• ./build.sh breakout --cpu-env: PASS
• ./build.sh breakout --profile: PASS
• ./puffer train breakout: reaches score 864; approximately 20-22M SPS on an RTX 3080 Ti depending on GPU clock state
• git diff --check: PASS

Notes

Experimental direct/parallel MinGRU scans, specialized small-Muon kernels, and larger-minibatch variants were profiled but intentionally excluded because they did not improve verified end-to-end training.