[flang] Make binary->decimal conversion buffer sizes accurate

The binary values that produce the most significant decimal

digits in an exact conversion are those with the least normal

biased exponent (1) and all fractional bits set, not the

least-valued subnormals. So the binary->decimal conversion

buffer sizes were a little short, and could cause a overrun crash.

Differential revision: https://reviews.llvm.org/D88688