In this release:

- fast-math is unconditionally no_std,

- approximations for `atan`, `atan2`, `exp` and `exp2`,

- `log2` is 1.25x faster in scalar code, and `log2_raw` can now be
  easily autovectorized, making it 5x faster (even with just SSE2 on
  x86-64), and

- `log2` on subnormal numbers had absolute error > 1 in some cases.