feat: Add usePlainNumberFormatForAllCells option to read numeric cells at full precision regardless of cell format

[vladislav-ishchenko]

Closes #1053.

What

New boolean read option usePlainNumberFormatForAllCells (default false): every non-date numeric cell read into a string column — including cached numeric formula results — is rendered through the existing PlainNumberFormat (full precision, no scientific notation), ignoring the cell's number format. Date-formatted cells keep their formatted rendering, text cells stay verbatim, non-finite values keep POI's display rendering. V2 column naming honors the option too, so a numeric header cell is named consistently with its own data cells.

usePlainNumberFormat only registers PlainNumberFormat for the General/@ format strings, so cells with an explicit number format still render their rounded/scientific display value: 84.789 under 0.00 reads as "84.79", and large formatted numbers read as scientific notation (#126, #771). Widening the registration isn't viable — the custom-format map is keyed by exact format string and shared with date rendering, and 2+-part ; conditional formats bypass the map entirely — so the format-independent path branches on the cell instead.

Also in this PR

PlainNumberFormat appended the unstripped BigDecimal, so single-significant-digit values below 1e-3 gained a spurious trailing zero from Double.toString's d.0E-x mantissa: 0.0005 read as "0.00050". It now appends the stripped value. Only that value class is affected (a sweep over 400k random doubles found no other differences), which means usePlainNumberFormat=true on a General-format cell holding 0.0005 now reads "0.0005" instead of "0.00050".

Tests

Both engines (V1 + V2), including the maxRowsInMemory streaming path: explicit-format rounding vs. plain rendering, General-format scientific notation, text cells verbatim, date cells, a cached numeric formula result, the 0.0005 trailing-zero case, numeric header naming, and a usePlainNumberFormat-only read pinning the boundary between the two options. The V1 tests go through the spark.read.excel(...) DSL so the option-key plumbing is exercised as well.