fix: support SAC multi-GPU replay IPC

[TATP-233]

Summary

• add a rank-local multi-GPU CPU-pinned replay pipeline for off-policy replay batches
• route per-rank replay pack request/ready queues and shared host slots through the collector pack service
• make the multi-GPU off-policy runner learner-class driven and use learner-provided initial parameter sync
• gate this validation round to SAC and add regression coverage for ranked replay IPC

Validation

• make test-all
• Remote 2-GPU smoke on A800 cards 0 and 7: CUDA_VISIBLE_DEVICES=0,7 uv run train --algo sac --task g1_walk_flat --sim mujoco training.num_gpus=2 algo.obs_normalization=false algo.use_symmetry=false completed successfully (FastSAC_x2GPU, 5000 iterations)