Fix integer overflow in LDA partition bookkeeping for large corpora

[sldyns]

Summary

This fixes integer overflow in the LDA multi-worker partition path for very large corpora.

Root cause

LDAModel::updatePartition() accumulated total token counts with a literal 0, which makes std::accumulate sum in 32-bit integer space before assigning back to a wider type.

For corpora whose total token count exceeds INT32_MAX, this can overflow and corrupt partition bookkeeping in the PARTITION path.

Changes

• use uint64_t for the partition total-count accumulator
• use uint64_t for the partition cumulative-count accumulator
• use uint64_t for totCf in the term-weighting path

Notes

This change is intended to preserve behavior for normal corpora while avoiding overflow for large-token inputs.

[sldyns]

I took a closer look at the failures.

Aside from macOS Intel + Python 3.9, the remaining failures all appear to be the same flaky reproducibility test (test_DTModel_reproducibility_NONE). The topic words and ordering are unchanged, and the diffs are only tiny floating-point differences (~1e-9), so this looks like an overly strict exact-equality assertion rather than a regression from this patch.

The macOS Intel + Python 3.9 failure is different (test_corpus_transform) and looks like a separate metadata/Unicode issue in the DMR/document binding path. This PR only changes src/TopicModel/LDAModel.hpp, so I do not think that failure is caused by this patch.