Add vecenv fallback and fix batched forward state

[ry2009]

Summary

• Add a safe vecenv fallback in create_environments: if breakout.so is missing symbols, we fall back to a minimal dummy VecEnv instead of crashing.
• Fix batched_forward state shape to {num_layers, minibatch_segments, 1, hidden_size}, matching PolicyMinGRU.
• Remove unused #include <stdatomic.h> in vecenv.h.

Why

• Previously, pufferlib.pufferl sps would segfault when the native env shared library wasn’t available or lacked OBS_N/ACT_N/OBS_T/ACT_T.
• The incorrect state shape in batched_forward could lead to shape mismatches for RNN policies.

Notes / Perf

• Fused CUDA kernels (including RMSNorm) build and run; on A100×2:
  • compile_puffer.py: ~7.35M inference SPS, ~2.45M train SPS (model-only microbench).
  • python -m pufferlib.pufferl sps puffer_nmmo3: runs without crashing; ~2.7M SPS with the dummy fallback.
• To report real env SPS, a native NMMO3 vecenv .so exporting OBS_N/ACT_N/OBS_T/ACT_T is still needed.