Enable metrics being computed even with NaN targets through the usage of masks

[giacomoguiduzzi]

1. Feature description

This feature would allow users to work on Irregularly Sampled Time Series when testing imputation methods by computing the loss directly through PyPOTS loss functions, which would make sense as the library is thought to handle missing data in Time Series.

2. Motivation

To evaluate an imputation model one possible solution is to predict a version of the validation dataset that has more missing values so that there is a "ground truth" to compute the metrics with; when working with Irregularly Sampled Time Series though, the validation dataset could have missing data. The loss functions from PyPOTS accept masks in input, but do not compute the metrics if NaN values are detected. It would be expected that the loss functions exploit these masks to ignore the NaN values.

3. Your contribution

PR #707 attempts to implement this feature by taking into account whether there are NaN values that are not covered by the masks, so as to correctly compute the metrics nonetheless.