Derive ShapedRun::advance from clusters

[waywardmonkeys]

process_clusters updated its run_advance accumulator only when it encountered the next cluster. It therefore omitted the final cluster, and a run containing one cluster reported zero advance.

Derive ShapedRun::advance from finalized ClusterData::advance values instead. This makes cluster data the source of truth and removes the parallel accumulator from process_clusters. Using finalized ClusterData::advance values as the source of truth also respects transformations applied while constructing clusters, such as stripping newline glyph contribution.

Add a regression test for a one-cluster run.

Performance

A repeated paired Tango comparison against fed192a found no statistically significant regressions:

Default Style - arabic 20 characters       -0.74%
Default Style - latin 20 characters        -0.06%
Default Style - japanese 20 characters     -0.42%
Default Style - arabic 1 paragraph         -0.60%
Default Style - latin 1 paragraph          +0.21%
Default Style - japanese 1 paragraph       -0.90%
Default Style - arabic 4 paragraph         -0.65%
Default Style - latin 4 paragraph          -0.44%
Default Style - japanese 4 paragraph       -1.10%*
Styled - arabic 20 characters              +0.22%
Styled - latin 20 characters               -0.74%
Styled - japanese 20 characters            +0.73%
Styled - arabic 1 paragraph                +0.44%
Styled - latin 1 paragraph                 +0.15%
Styled - japanese 1 paragraph              +0.30%
Styled - arabic 4 paragraph                +0.52%
Styled - latin 4 paragraph                 +0.68%
Styled - japanese 4 paragraph              -0.94%

[waywardmonkeys]

An LLM (GPT 5.6, Sol, xhigh) was involved in this with finding / identifying the issue and writing the test.

[tomcur]

For additional context, see: #679 (comment).

[taj-p]

Nice!! Thank you 🙏