The score minimises to zero when the observation equals the estimated
quantile, so that smaller scores indicate a better fitting model.
Positive residuals are penalised by a factor of tau, and negative
residuals by tau - 1. This loss is also known as the stick function, check
loss, asymmetric absolute deviation, or pinball loss.
For observation x, estimate x_hat, and level tau, the score
(c.f. Gneiting, 2011) is
$$
S_\tau(x, \hat{x}) =
\begin{cases}
\tau |x - \hat{x})|, & x \ge \hat{x}, \\
(1 - \tau)|x - \hat{x}|, & x < \hat{x}.
\end{cases}
$$
Vector recycling of all three arguments follows the rules in
vctrs::vec_recycle_common().