The generic WG size binary was broken for the CPU devices and the
brokenness typically is hidden by the PoCL's feature of (trying to)
specialize also WGs loaded from binaries per local size. PoCL just
generated always a new WG function and never used the generic WG
function by default if online compilation is enabled in the build.

Now this is tested automatically by adding a (undocumented for the time
being) config env POCL_WORK_GROUP_SPECIALIZATION, which when set to 0
brute force disables specialization to test the binary loading and the
generic local size WG function generation.