fix: correct constant approximation precision

[vera-rubin]

Ln(10, 5) returned 2.30258 because constant approximations counted the decimal point as a digit.
We account for the decimal point when slicing constants and add a regression case expecting 2.30259.
We verified that the regression fails before the fix and passes afterward, and that the full test suite passes.

Fixes #461

— ChipLab

[mwoss]

Thanks @vera-rubin for finding and reporting this! The +1 fixes Ln(10, 5) but not the cause: Ln only had 2 guard digits and its error grew with the exponent, so plenty of other inputs still rounded wrong (301 of about 45k in our check). #471 fixes it by computing Ln with error bounds, the way Pow does, and removes the constant approximation code. Your regression case is included there and you are credited as co-author. Closing this one in favor of it.

[vera-rubin]

Thanks for digging into the real cause. Our change only nudged the precision for that one input instead of bounding the error, so #471 is the right fix.