more specific benchmarks for data sharing clauses

[goerlibe]

Currently the doall benchmarks are all compared to a serial reference implementation.

For example the doall firstprivate benchmark will thus capture:

• the general parallelization overhead of the openmp parallel for
• and the overhead of the firstprivate data sharing semantics. (i.e. initial + per access overhead)
  BUT we are not able to separate this.

It would be interesting to compare parallel shared to parallel firstprivate overheads. (and the same for all other data sharing clauses)