Investigate faster performance than upstream implementation

As observed in https://github.com/dtolnay/ryu/pull/34#issue-496690273, the 32-bit and 64-bit float-to-string benchmark in this repo is something like 40% faster than the exact same benchmark in https://github.com/ulfjack/ryu. We need to examine whether there is any inadvertent discrepancy between the two benchmarks, difference in build configuration, or any toolchain difference that would account for this. The code is a line-by-line transliteration so there should really be no performance difference, and certainly not this big.