[autoparallel]add embedding handler

[YuliangLiu0306]

What does this PR do

1. This PR introduces embedding handler to generator valid strategies for embedding handler.
2. Construct unit test to check the numerical correctness.

What need we do in future

1. The computation cost is estimated as a dense operation, instead of a sparse one. Therefore, we need to take care of it when we want to get the correct time cost estimation.
2. The weight of embedding has 2 dimensions, one is num_embeddings, another is embedding_dims. In this PR, only embedding_dims get shard, because num_embeddings case will require more actions in runtime preparation passes.

[YuliangLiu0306]

I combine the last dims and batch dimension together, instead of doing these actions one by one. Consider this example:
tensor shape: (1, 4, 16)
Device mesh: (2, 2)
ShardingSpec before converting: S0S1
If we recover the last dimensions first, the new dim_partition_dict will be {0: [0], 2: [1]}, 0 is inherit from the old_partition_dict({0: [0], 1: [1]}) and 2 comes from the dim_mapping {1: 2}. Then, the new ShardingSpec initialization process will crash, because the size of 0 dimension is 1 which is not dividable by 2.