[FEATURE]: Implement LayoutConverter

[YuliangLiu0306]

Describe the feature

LayoutConverter mainly has three main functions:

1. Search an efficient transform path for the given pair of source_layout and target_layout.
2. Estimate the converting cost for the given pair of source_layout and target_layout.
3. Apply the layout converting path to redistribute a tensor from source_layout to target_layout.

To search the transform path, we use a heuristic algorithm which could be simplily describe as:

• Step1: Generate all one-step transform sequences from source_layout.
• Step2: Pick the 'best' layout following the heuristic function.
• Step3: Repeat above steps until the source layout transform to target layout.

To estimate the total converting cost, we implement a method to estimate the communication cost of each communication operation under the instruction of CommSpec. Then, we just need sum the costs of CommSpec in the transform_sequence up to get the total converting cost.