Dev/manage meta

[BeachWang]

Have organized and consolidated the generated metadata uniformly.

1. Meta data produced by mapper OPs is stored in Fields.meta field. Corresponding keys are list in MetaKeys.
2. For convenience of split a batch sample, meta data produced by aggregator OPs is stored in Fields.agg field. Corresponding keys are list in AggKeys. Here, we add a default Fields.agg field to all samples in all aggregator operators, without adding a decorator marker TAGGING_OPS like we do with mappers.
3. Unify the interface for generating meta operators, allowing modification of meta keys. If a meta key already exists, it should simply return without making any changes.
4. Add corresponding unit test.
5. Adjust the role playing demo according to the changes.
6. Remove nested_set function.

[yxdyc]

How about removing Fields.agg and AggKeys? Instead, we can unify them into Fields.meta and MetaKeys. This makes sense because they are implemented in a very similar way, and conceptually, agg is still a type of meta.

Additionally, currently, we only have two types of AggKeys, and keeping both agg and meta may increase the cognitive load for users and introduce inconsistencies. Our 2.0 paper mentions that we have three main parts of the fields: 1. data payload, 2. stats, 3. meta. Unifying agg into meta would align better with this structure.

Others LGTM.