Fix cast to Python scalar in Matthews Correlation metric

[mariosasko]

This PR is motivated by issue #2964.

The Matthews Correlation metric relies on sklearn's matthews_corrcoef function to compute the result. This function returns either float or np.float64 (see the source). Obviously, calling .item() on the float value will fail, so I'm fixing this with the built-in float() function, which covers both cases. Surprisingly, on my machine, casting np.float64 to a Python scalar with float() is even faster than with the .item() method.

[lhoestq]

Cool thanks :)