Estimate redistribute_cost by _TransformInfo

[mori360]

Estimate redistribute_cost by _TransformInfo rather than comparing source and destination state
Here are the changes:

1. use _TransformInfo to collect shard order
2. use _TransformInfo to explore the path during redistribute. The previous approach only considers the source and destination state and a one-step redistribute for each placement combination, however, it could be incorrect in some cases:
   a. S(0)S(0) -> S(0)R, need 1 allgather
   b. S(0)S(0) -> RS(0), need 2 allgather, which could not be found if only care S(0)->R
3. use _TransformInfo.logical_shape to estimate comm_byte. The current comm_bytes_gb is based on tensor shape and number of shards. In case 2.b, the comm_byte for 2 allgather is different.

TODO:
There are some compute_cost with comm_bytes_gb, need to verify whether they could return the expected cost.

[fmassa]

In general, what we would like to have here I think is the minimal redistribution cost over all possible input/output orderings.

This is to ensure that we don't have to increase the search space for AutoParallel when performing the optimization, as we can focus only on the shardings (without order) and then optimize the ordering afterwards.

Does it make sense?