Replacement for simd + simd_opt. Builds on stable rustc 1.27+

The coresimd feature overrides all other simd flags. Build with coresimd
will fail for unsupported target platforms, e.g. non-x86.

Benched on a Xeon X5650:
sse2 only: faster than simd_opt
ssse3: slightly faster than simd_opt

Benched on a Haswell:
avx2: same speed as simd_opt

This will test the sse2 and probably ssse3 coresimd implementations. If
the buildbox happens to support avx2, it will also test the avx2
implementations.