This reverts commit 9c2d4d57.
(In fact, it only reverts the important hunk)

   * Worker threads (not master) will start tagged (_mustSleep) until
      they're added to a team
   * Worker threads (not master) that leave a team will be also tagged

   * OpenMP apps can now modify their number of threads using
      nanos_omp_*_mask()

Fixes a problem of types in CachedAccelerator, which solves the issue of running cholesky with big problem sizes in some systems (closes #812)

_targetThreads is initialised before loadNUMAInfo, although the final value will be computed afterwards. You will still see warnings in some cases (GPU, OpenCL), but it will be like "adjusting from 12 to 14". Fixes #809.

New teams with higher number of threads were not notifying the new
size of reserved threads to Extrae.

Fixed a bug introduced in [9d72d14b] that caused the System initialisation not to wait for the creation of CL and GPU threads

Fixes problems with the order of NUMA information reading, PE creation and module creation (fixes #804, fixes #806)

Adding request to (plain) dependencies domain system to check if there are any pendant output for a given memory address; see #802

Changing config options names for work first scheduler, changin default values to fit Cilk scheduler; see #805

   - Master thread is added to the list.
   - stopAll now stops also worker threads from PE 0.

Adding stub service on dependences API to request if there are pendant writes on a given address; see #802

It fixes  #801