Team serialization may fail when the program exits abnormally

[Terrdi]

Bug description
When I tried to run the Werewolf example, the model call failed due to platform flow restrictions. At this time, the program failed to serialize when it tried to write the current state to the storage/team.json file.
After debugging, I found that the attribute special_actions in BasePlayer contained Action classes such as Speak, which caused the serialization failure

Bug solved method

I tried to add a custom serializer to Team so that it can be customized by users
#1688

Environment information

• LLM type and model name: deepseek-r1
• System version: MacOS
• Python version: python3.9
• MetaGPT version or branch: main

• packages version:
• installation method: pip install -e ".[rag,test,dev]"

[seehi]

The PR is large.

[Terrdi]

OK, I changed the target branch and submitted some new changes.
The aflow example also uses json serialization to save the results. The total_cost field is numpy.int64, which cannot be serialized directly. Then I fixed this problem.