Add vecwid256 microarchitecture option to enable -mprefer-vector-width=256

This defines a fake vecwid256 microarchitecture flag that expands to -mprefer-vector-width=256.

It does not otherwise seem to be possible to define flags containing - and = via the $BINARY_TAG variable.

With this change then skylake_avx512+vecwid256-centos7-clang8-opt can be used to build with AVX512 instructions with 256bit wide registers.