Deferred minor findings from the archive and OpenDocument reviews

[vaceslav]

Minor findings of the final reviews of #19 (zip archives) and #13 (OpenDocument). They were deferred on purpose: none of them gives wrong data on files a real writer produces. They are listed here so they are not lost. Each item says what a user would see.

Archive (ArchiveCursor, ChunkedBuffer)

• A small workbook inside an archive costs a 1 MB allocation, twice (once when the sheets are listed, once when they are read). When the first piece fills exactly to the declared size, the loop allocates a whole piece just to find the end of the file. The remark saying "a small workbook does not cost a megabyte" is therefore wrong. Fix: probe with a one-byte read before allocating the next piece.
• ReadRow after a failed reopen throws NullReferenceException, not a TabularException. Only running out of memory reaches this. Separately, CurrentSheetIndex is left stale when the inner MoveToSheet throws.
• The archive-wide sheet ceiling reuses XlsxCursorOptions.MaxSheets for csv and ods sources as well, and reports the limit as the bare string "MaxSheets". Fix: document which option applies, or use nameof.
• Two entries with the same path produce two sheets nothing tells apart: same name, same Source. Fix: skip the later duplicate, or document it.
• TabularFile.ClassifyZip reads a zip's central directory once more before any entry-count bound applies. This is negligible for real files, but doubles the cost for a hostile directory. Fix: read the entry count from the end-of-central-directory record first.
• Polish: ArchiveCursor.Dispose has no <inheritdoc/>; ChunkedBuffer throws ArgumentOutOfRangeException rather than IOException on a negative seek, and returns 0 when read after dispose.

OpenDocument (OdsCursor, TableNameScan)

• A paragraph nested inside a paragraph drops the outer one's tail. <text:p>ab<text:p>in</text:p>ef</text:p> reads ab\nin. The markup is malformed ODF.
• UTF-16 XML is not refused as XML. A .fods or Excel 2003 XML file in UTF-16 with a BOM gets past IsXmlDocument and is read as csv. Fix: also check FF FE/FE FF followed by <?xml.
• NaN, Infinity and 1e400 in office:value become Number cells. The xlsx reader does the same, so the two agree. A double.IsFinite guard in both would keep such values away from analysis.
• Every time cell allocates a string (XmlConvert.ToTimeSpan takes a string), and a malformed time-value costs a thrown and caught exception, about 22 µs per cell. A small span-based ISO-duration parser would remove both.
• In ODS mode, the error for repeated attributes still says "repeats its r, t or s attribute", which is the xlsx wording.
• When the OdsCursor constructor fails after the content part is opened, its scanner is not disposed. The archive underneath is disposed, so only a pooled buffer leaks.
• table:null-date in table:calculation-settings is ignored when time durations are read. Durations always count from the 1899 null date.
• OdsCursor.ReadRow does not wrap InvalidDataException as format.corrupt, as XlsxCursor does. It is hardly reachable, because the name pass inflates the whole part first.
• The table-tag ceiling in TableNameScan, and the scanner's 16 M-character node ceiling, ignore a configured MaxValueChars. Sheet names have no counterpart of xlsx's MaxMetadataChars.

Related, but an API change and so tracked in #15: ITabularCursor.MoveToSheet takes no cancellation token (see KNOWN-ISSUES).