Add multiprocess sim environment

[kyhavlov]

Adds a multiprocessing env wrapper for the JAX sim rollout worker to use. The parent process keeps policy inference and trajectory assembly, and worker processes each own a shard of SimBatchedEnvironment and write observations/resets into shared memory. Uses the same num_envs / inner_batch_size = worker count logic as the dolphin multiprocessing env code, and is enabled by async_envs=True. Doesn't support num_env_steps for now.

benchmarking in run_evaluator.py:

Settings:
4096 envs, rollout_length=3000, inner_batch_size=256

Non-MP:
timings: {'env_push': '27.776', 'agent_step': '5.610', 'trajectory_build': '7.411'}
env_fps: 97938.49, player_fps: 195876.99, sps: 23.91

MP 16 x 256:
timings: {'env_push': '4.066', 'agent_step': '5.470', 'trajectory_build': '12.154'}
env_fps: 179583.63, player_fps: 359167.25, sps: 43.84

Also verified a 100-step sim RL test run using the multiprocess env still trains - the trained checkpoint beat the base checkpoint 92-8 over 100 sim games.

[kyhavlov]

Also worth noting: this now writes sim observations straight into the rollout trajectory buffer. GameBatchBuffers.time_major(...) owns the full T+1 observation tree, and each timestep has a GameBatchBuffers view that the env can fill directly. That removes the old extra copy from “current env observation” into the trajectory buffer.

same benchmark timings after removing the extra copy:

16 workers x 256 envs:

timings: {'env_pop': '0.000', 'env_push': '3.886', 'agent_step': {1: '1.808',
2: '1.808'}, 'trajectory_build': '8.091'}
env_fps: 247455.60, player_fps: 494911.19, sps: 60.41

[vladfi1]

This appears to be unused.