fix: Use a more flexible approach to check favorability of metrics

[glemaitre]

closes #1061

We are more flexible using a regular expression to check the score names. In addition, we take care to test first the neg_ part that would mean that negative score are therefore "higher is greater" convention.

[augustebaum]

The "neg_ means higher is better" rule makes me a bit wary of false positives.
For example, neg_accuracy gets falsely considered as higher-is-better.

[glemaitre]

It will never happen in practice. The neg_ is only a scikit-learn convention used to negate the loss/error/deviance such that whatever metrics used in a grid-search will be maximized.

So if you are a user, you are never going to defined your own neg_ metric to make a score that lower is better.

[augustebaum]

Sounds good.

[augustebaum]

Thanks!