Refactor UTF-8 decoding to enable use of SIMD.

[christianparpart]

Implements #20.

CHECKLIST

• basic SSE2 accelerated decoder (128-bit version; based on BobSteagall's utf_utils)
• AVX decoder (256-bit?)
• AVX2 decoer (256-bit)
• AVX512 decoder?
• AVX512-BW accelerated decoder (512-bit version)
• only compile SIMD variants if target arch does support it.
• default to optimal implementation for target host
• add accelerator::runtime_detect to chose best implementation at runtime (the result should be cached but also the cache should be invalidatable?)
• improve API documentation on members

[data-man]

add accelerator::runtime_detect to chose best implementation at runtime

Can we use some third-party library for this (and in Contour too)?
E.g. https://github.com/google/cpu_features, https://github.com/anrieff/libcpuid or https://github.com/steinwurf/cpuid.

[christianparpart]

add accelerator::runtime_detect to chose best implementation at runtime

Can we use some third-party library for this (and in Contour too)?

I think CPU feature testing is so simple and just a few lines of code with regards to what we need. There is no need to pull in another dependency. :)

p.s. i have the code already around.

[christianparpart]

outdated, and most importantly, superseeded by our new scan API.