[skip lazy allocation feature] add comm part to et_replay without lazy allocation of tensors

[TaekyungHeo]

Test Plan

#!/bin/bash
#SBATCH --job-name=full_replay_job
#SBATCH --nodes=1
#SBATCH --ntasks=8
#SBATCH --ntasks-per-node=8
#SBATCH --gpus-per-node=8
#SBATCH --time=01:00:00
#SBATCH --output=%x_%j.out
#SBATCH --error=%x_%j.err


srun -A general_cs_plutus\
     --container-image /lustre/fsw/portfolios/general/users/songyant/docker/songyant+docker-images+ai_develop.sqsh \
     --no-container-mount-home \
     --container-mounts=/lustre/fsw/portfolios/general/users/songyant/workspace_local:/workspace_local \
     zsh -c "

        if [ ! -d /run/sshd ]; then
            mkdir -p /run/sshd
            chmod 0755 /run/sshd
        fi

        source ~/.zshrc
        conda activate full_replay

        et_replay --trace_path /workspace_local/full_replay/traces/resnet-8gpu
     "