a preliminary implementation of single-process simulator, which still has problems

[Corle-hyz]

Support to simulate a single rank of distributed training. Rank 0 is allowed now. Also, this PR includes a preliminary implementation of auto-tuning in cooperation with the simulator.

INFO: This simulator will not normally exit, so it must be killed manually. Further improvement is needed.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}