[autoparallel] add resnet autoparallel unit test

[YuliangLiu0306]

This PR does two things:

1. Add resnet autoparallel unit test.
2. Add the weight grad communication cost for conv. Typically, the weight of convulotion network is much smaller than activation and the all reduce communication could be overlapped with the backward computing, so it is not counted in previous PR. However, in some extreme cases, the parameters could be much larger than activation, if the communication cost to computing backward parameter grad is still omitted, solver could not get a reasonable solution.