[Experimental] Update to a better heterogenous impl

[aoyulong]

This pull request is for re-implementing heterogeneous training in a better way. The goals are as follows:

• Supporting more than 2 hardware devices
• Allowing the size of tensor parallel, context parallel, and pipeline parallel to be different
• Decoupling some code from the original implementation

Please note that the code is in a very early stage and there is no guarantee that it will work well.