fix(stats): handle empty documents - #512

Summary

• return 0.0 for ratios and averages whose denominator is empty
• apply the same behavior across document, line, paragraph, sentence, word, and word-contamination stats
• add regression coverage for empty documents across all affected stats blocks

Empty or fully filtered documents currently raise ZeroDivisionError in several
stats blocks. This can stop a data-audit pipeline before downstream cleaning has
a chance to handle the sample. The helper preserves existing results for
non-empty inputs and keeps count metrics such as n_lines unchanged.

Tests

• focused stats regression tests — 7 passed, 6 subtests passed
• Ruff lint and format checks
• git diff --check