Cache with running average - possible bias and instability

Our cache uses a running average to speed up the calculation of the shapley values for non-deterministic training. However, it is unclear how stable montecarlo shapley is towards non-deterministic training, and even more so with a running average calculation of the shapley values. We should do some experiments to check the stability of these new options in a few examples, and possibly include some tests in our pipelines.

该 issue 涉及使用运行平均值缓存进行非确定性训练时 Monte Carlo Shapley 值计算的稳定性。首先检查 pyDVL 中的缓存实现和 Shapley 值计算代码。设计实验，在不同数据集或模型上比较使用缓存和不使用缓存的结果，以评估偏差和不稳定性。最终应形成一份报告或测试套件，用于验证该方法。

由索引模型根据 Issue 内容生成。