[autoparallel]add backward cost info into strategies

[YuliangLiu0306]

This PR add backward information into conv strategies, the information including:

1. memory cost: I use a memory pair to store the memory cost, first element of pair is forward memory cost and second one is backward memory cost which is composed with activation grad cost and parameters grad cost.
2. computing cost: compute cost of backward has two parts, first is activation computing cost and second is parameter computing cost.
3. communication cost: In previous, we only count the communication cost during forward. In the PR, the communication cost of backward is counted. Additionally, in some no forward communication cost case, the communication may happen during backward phase.
4. resharding cost: resharding cost will also happen during backward, such as we convert the input sharding spec from [R, S] into [R, R], which means we need to recover it from [R, R] back to [R, S] during backward phase.