Refactor source manager for policy, reference and reward model

# Description

For now, the resource for policy is computed from the whole resource, inference resource and rewaed model resource. This leads to the inability to customize the training resources.
For example, in non-colocated condition, if there are 2 nodes and 8 gpus per nodes, 4 gpus for inference, 4 gpus for rm env. Then the resource for policy can be 8 gpus on 1 node, and also can be 2 nodes and each with 4 gpus.

# Refactor

Modify config, and specified resource for each actor. Set up a separate resource management module.