[atari] Fix frame_skip=1 max-pool ghosting

[Trinkle23897]

Summary

• Problem: frame_skip=1 Atari steps were max-pooling the current frame with stale data from maxpool_buf_[0], which leaves bright ball/paddle trails in RGB observations.
• Scope: Fix the Step() frame-buffer indexing and skip max-pool when there is only one frame to pool; no intended behavior change for frame_skip>1.
• Outcome: frame_skip=1 now emits the current frame without stale-frame ghosting, while frame_skip>=2 keeps the existing max-pool behavior.

This is a minimal Atari-only fix for the one-frame max-pool edge case.

Technical Details

• Approach: Use min(frame_skip_, 2) as the number of pooled frames, write the sole frame into maxpool_buf_[0] when frame_skip=1, and only call PushStack(..., maxpool=true) when frame_skip_ > 1.
• Code pointers: envpool/atari/atari_env.h
• Notes: Reset-time stack fill is unchanged.

Test Plan

Automated

• git diff --check: passed
• make bazel-pip-requirement-dev && USE_BAZEL_VERSION=8.6.0 BAZEL_RULES_QT_DIR=/opt/homebrew/opt/qt@5 bazelisk test --test_output=errors --config=test --spawn_strategy=local --action_env=PATH=/Users/jiayi/.virtualenvs/openai/lib/python3.12/site-packages/cmake/data/bin:/Users/jiayi/.virtualenvs/openai/bin:/opt/homebrew/bin:/bin:/usr/bin:/usr/local/bin //envpool/atari:atari_env_test: passed

Suggested Manual

• Run a frame_skip=1 Atari RGB rollout and inspect a few frames/video to verify the ball and paddle no longer leave one-step trails.