Since i suspect the bash overhead to be the reason for the overhead
observed in musl libc that made clang-hash slower than recompilation.

Too fool ccache, we add a spurious semi-colon in the global space. To
avoid warnings, which could slow down compilation, we have to suppress
the warnings and ignore the empty declaration in the visitor. The
warnings are only visible with -pedantic, which is no problem for the
data we have gathered so far.

ccache does not redo the compilation, if only a newline was added. But
a single semi-colon does do the trick.

Since we want to measure an end to end time, we have to use the
bare-mininum wrappers that only call clang with {nothing, ccache, clang-hash}

The changes test cases were all brittle and fragile regarding the system
configuration. On Linux, they actually were correct, but only on pure luck.

This change speeds up the compilation of the plugin significantly.