Fix brier_class() case weight normalization

[taekop]

Closes #633.

brier_ind() normalized case_weights with a softmax (exp(w - max(w)) / sum(exp(w - max(w)))) before computing the weighted mean. This is only equivalent to a plain weighted mean when all the weights are equal; once weights are spread out, exp() distorts their relative sizes and can underflow small weights to zero, silently dropping those observations. The fix removes the softmax and normalizes by the plain sum of weights instead, matching how other weighted metrics like mn_log_loss() handle case_weights.

Since ranked_prob_score() calls the same brier_ind() helper, its weighted calculations were affected too and are fixed by the same change; I updated its case-weight test's expected value, which had been generated from the buggy output rather than the stated orf::rps() reference.

Added a test comparing brier_class() with widely spread case_weights against the same data expanded with tidyr::uncount().