fix(analysis): break ties between number readings by the sheet's evidence, and flag disagreement

[vaceslav]

Closes #61. This is the numeric twin of #58.

A column of values with exactly three digits after one separator (48.137) reads completely both as decimals and as German grouped integers. The narrowest type used to win the tie (Integer/de-DE, 48137), while MinNumeric/MaxNumeric were measured under the first culture listed (48.137). The top hypothesis and the extremes described different numbers.

Changes

• New fact ColumnFacts.NumberReadingsDisagree. It is true when the cultures that read the most values as numbers read some row as different numbers. It is measured with the same per-row, order-free fingerprint as DateReadingsDisagree.
• Sheet-level evidence for the decimal separator, in this order:
  • the sheet's columns whose values are not in doubt, meaning a comma culture reads more of them as numbers than a point culture does, or the other way round;
  • a csv's delimiter: ; implies a comma, and , implies a point, because a comma-delimited file cannot use an unquoted comma for decimals.
• Ranking, only where the readings disagree: the favoured separator's readings come first, and a decimal ranks above a grouped integer, because one group is weak evidence of grouping. Where the readings agree, the order is unchanged.
• Extremes: MinNumeric/MaxNumeric are measured under the same choice, so they describe the reading that ranks first.

This also covers the 41,181 observation in #47.

Docs: docs/concepts.md gains "Which columns a sheet has", which notes that a value to the right of the header row makes a column with an empty header. It also gains "When readings tie", which explains the tie rules and both flags, and that a leading group may be any length (1234.567 under de-DE is 1234567).

Performance: on the corpus, within −3.2 % to +2.3 % of main (5M-row csv +2.3 %). The fingerprint hashes the decimal's raw bits, and skips a culture that reads numbers exactly as an earlier one does.

Tests: AmbiguousNumberTests, six cases: a comma-delimited csv, a semicolon-delimited csv, evidence from another column in both directions, xlsx text cells with no evidence, and columns where the readings agree.