[lazyinit] combine lazy tensor with dtensor

[ver217]

Overview

We have implemented a single-process version. We may want lazy tensor can be distributed during/after materialization, this feature may be powered by dtensor.

Wanna track the development progress? Take a look at

proposal: #3124
kanban: Lazy initialization

Goal

We have to find out an acceptable solution for both auto-parallelism and gemini.