Add Craftax moonshot environment

[Infatoshi]

Summary

Adds a side-by-side craftax_moonshot implementation for the 4.0 branch. This keeps the existing Craftax implementation intact while adding a CPU-oriented version with a smaller packed float observation path, reset-state caching, tiled stepping, and a per-buffer VecEnv rollout hook used only by environments that opt in.

The src/vecenv.h change is intentionally narrow: default environments keep the existing c_step rollout loop, while craftax_moonshot defines MY_VEC_STEP_RANGE to avoid the scalar fallback during Puffer rollouts.

Measured throughput

On a Ryzen 9 9950X3D + RTX PRO 6000 Blackwell workstation, pinned to the V-Cache cores:

env OMP_NUM_THREADS=1 taskset -c 0-7,16-23 \
  uv run puffer train craftax_moonshot \
  --train.total-timesteps 200000000 \
  --vec.total-agents 16384 \
  --vec.num-buffers 16 \
  --vec.num-threads 16 \
  --policy.hidden-size 32 \
  --policy.num-layers 1 \
  --policy.expansion-factor 1 \
  --train.horizon 128 \
  --train.minibatch-size 32768

Short-run dashboard measurement: about 4.0M training SPS with env around 47% of wall time.

Raw random-action env-only measurement on 4096 agents: about 16.8M env SPS.

Verification

uv run --with pybind11 --with rich_argparse --with nvidia-cudnn-cu12 --with nvidia-nccl-cu12 ./build.sh craftax_moonshot
uv run --with pytest pytest tests/craftax_diff.py -q
uv run --with ruff ruff check tests/craftax_diff.py --fix

Results:

Built: pufferlib/_C.cpython-313-x86_64-linux-gnu.so
2 passed in 15.07s
All checks passed!

Repo-wide pytest/ruff are currently blocked by pre-existing upstream dependency/lint issues unrelated to this PR.

[Infatoshi]

Closing this version. Reworking the patch to update the existing craftax environment directly instead of adding craftax_moonshot side-by-side.