Add quantile metric.

[trivialfis]

At the moment, the metric doesn't return the quantile value as part of its name like [email protected] (metrics like NDCG would display ndcg@3). It's just quantile. Eventually, we need a better interface for metrics to support non-trivial distance measures. This PR only handles the basic implementation of mean pinball loss.

• Add quantile loss as metric, currently simply named quantile. We can call it something more "formal" like mean_pinball_error if it's deemed better.
• Add quantile parameter, which requires parsing vector of floats. We have similar structures for tree param with mono and interaction constraints, but for integer instead.

Extracted from #8750 .

[RAMitchell]

Can we do a small ground truth test against sklearn?

[trivialfis]

@RAMitchell Added a test with sklearn mean_pinball_loss.

[trivialfis]

Added a fix to parameter validation.