• On ARM64 (AArch64), replaced the inline assembly implementation of
  128-bit multiplication with a __uint128_t-based version.
  This avoids potential miscompilations of the experimental MuseAir
  hash observed with aggressive compiler optimizations (e.g. -O2/-O3)
  on some toolchains (notably Apple Silicon), while still generating
  optimal code on modern compilers.
• Fixed the speed computation of the undocumented -T option for the
  recovery functions. The reported values are now comparable to those
  of the generation functions. Note that the real speed is unchanged
  from the previous version, only the way it is measured has changed.