candidate-pipeline: sink non-finite ranking scores

[thepriben]

Summary

partial_cmp returns None for NaN. The selector treated that as Equal, so a candidate with an invalid score could stay ahead of a valid one and consume a Top-K slot.

Concrete case, K = 2:

input:     NaN    0.90    0.80
old sort:  NaN    0.90    0.80
selected:  NaN    0.90          // 0.80 dropped

After this change, non-finite scores (NaN, +inf, -inf) rank below every finite score. Finite scores still sort descending:

input:     NaN    0.90    0.80
new sort:  0.90   0.80    NaN
selected:  0.90   0.80

Missing scores (None) were already mapped to -inf. This only changes Some(NaN) / Some(±inf).