Issue with metrics of difficulty_mask_ratio and entropy

Hi team,
- I've identified an issue where the difficulty_mask_ratio metric gets overwritten during multi-domain training, causing us to lose valuable per-domain information.
In RLVRPipeline, the code calculates the difficulty_mask_ratio for each domain within a loop.
https://github.com/alibaba/ROLL/blob/9b85e63ad4c715aa6602ba2a43657e25217c7732/roll/pipeline/rlvr/rlvr_pipeline.py#L419
In this function, this ratio is then logged to the static metric key actor/difficulty_mask_ratio.
https://github.com/alibaba/ROLL/blob/9b85e63ad4c715aa6602ba2a43657e25217c7732/roll/utils/functionals.py#L608
Because the same metric key is used for every domain, the value is overwritten in each iteration. As a result, we can only see the difficulty_mask_ratio of the very last domain processed in the logs. I suggest making the metric key domain-specific.

- Also, the entropy is calculated at once https://github.com/alibaba/ROLL/blob/9b85e63ad4c715aa6602ba2a43657e25217c7732/roll/pipeline/rlvr/rlvr_pipeline.py#L401
but it assign to every domain's metrics with the same value, how about calculate every domain's entropy dividely?

Start in roll/pipeline/rlvr/rlvr_pipeline.py around lines 401 and 419, then inspect roll/utils/functionals.py around line 608. Trace how entropy and difficulty_mask_ratio are assigned inside the per-domain loop. Done means logs retain separate difficulty_mask_ratio values for each domain and entropy is calculated and reported per domain.

Written by the indexing model from the issue text.