[TENSOR] eager mode support about mars

[qinxuye]

I think in order to reach the availability of eager mode, we can separate the develop process into a few steps.

• Record which tensor has been executed in a specified session. (#80)
• Update the shape and so forth after execute.(#120)
• Handle the executed tensor by replacing it with tensor whose op is Fetch. Handling logic for local, local cluster and distributed session may be different.(#180)
• Provide a base class for the TensorOperandMixin which has a new_tensors method, maybe we need to implement a new_entities for the base class, and call it inside the new_tensors. This is a preparation for the coming dataframe module. (#173)
• If eager mode is on, make the new_tensor(maybe new_entity) trigger the execution when building expressions, we need to take care of the case cautiously, because new_tensor is also widely used in the tile, we need to make sure that the new_tensor in tile would not trigger the execution.(#200)

cc @NJUPole if there is any additional suggestion.

[qinxuye]

As all steps have been achieved, this summary issue can be closed.