Add tables of small powers of ten used in the fast path. The tables are redundant: We could also use the big, more accurate table and round the value to the correct type (in fact we did just that before this commit). However, the rounding is extra work and slows down the fast path.
Because only very small exponents enter the fast path, the table and thus the space overhead is negligible. Speed-wise, this is a clear win on a [benchmark] comparing the fast path to a naive, hand-optimized, inaccurate algorithm. Specifically, this change narrows the gap from a roughly 5x difference to a roughly 3.4x difference.
[benchmark]: https://gist.github.com/Veedrac/dbb0c07994bc7882098e
The import has been unnecessarily complicated since ParseFloatError::Invalid is not longer used unqualified.
The pfe_* functions do not need to be public any more since the only other use site, from_str_radix for floats, has been removed.
Move private bignum module to core::num, because it is not only used in flt2dec.
Extract private 80-bit soft-float into new core::num module for the same reason.
Stop using stability to hide the implementation details of ParseFloatError and
instead move the error type into the `dec2flt` module. Also move the
implementation blocks of `FromStr for f{32,64}` into `dec2flt` directly.
This commit primarily adds implementations of the algorithms from William
Clinger's paper "How to Read Floating Point Numbers Accurately". It also
includes a lot of infrastructure necessary for those algorithms, and some
unit tests.
Since these algorithms reject a few (extreme) inputs that were previously
accepted, this could be seen as a [breaking-change]
