dec2flt is slower than it should be

[rkruppe]

To some degree, slowness is unavoidable because correctly parsing floats can get pretty hard. However, @Veedrac has recently demonstrated that a simple, obvious, not-correct-in-general implementation is easily 5x faster than the current "fast path". I am certain this can be improved, if not to equal performance then at least to something between 1x and 2x.

On the other end of the spectrum, the slow path is very slow indeed. Here it's less a question of micro optimization and more of smart algorithms. The glibc approach sounds promising and would also solve a long-standing issue of not being able to handle some pathological inputs. There's no better implementation to benchmark against, other than perhaps libc::strtod, so I have no idea how much room for improvement there is and when this issue could be considered "closed". We'll see.