Add `quantile_score()` which accepts user-defined quantiles

[robjhyndman]

1. This function is largely redundant as it averages over the 99 percentiles, so it gives a value that is half of CRPS. It would be more useful to have percentile scores for a user-defined list of percentiles, and the function return a vector corresponding to those percentiles.
2. I think the pinball loss should be defined as exactly twice what is currently coded. That way, the loss at 50% equals absolute error and the average loss equals CRPS.

[mitchelloharawild]

Do you think percentile_score() should have a different default percentiles, in addition to being user defined?

[robjhyndman]

Maybe. Returning a vector of 99 percentiles makes intuitive sense given the name but is a little unwieldy. We could rename it to quantile_score with default prob = c(0.05,0.25,0.5,0.75,0.95).

[mitchelloharawild]

I would prefer to keep percentile_score() and have it call quantile_score() for backwards compatibility with minimal cost. We can encourage use of quantile_score() instead.

[mitchelloharawild]

The definition of pinball loss or quantile score?
I've changed pinball loss here: 26c689c

[robjhyndman]

The definition of quantile_score to match pinball_loss.

[mitchelloharawild]

Done.