If 'fib' is a private variable, compilers with a high enough optimization
level can detect that 'computation' is dead code and optimize too much.

* Fix old API names
* Fix LB cast to double
* Increase default loads and parallel granularity
* Improve MPI+OpenMP test:
  - Use performance libs instead of debug
  - Set some OMP env. vars to improve performance in DLB
  - Document the MPI no-binding flag

* Add verbose messages during init
* Finalize DLB if OMPT failed

Add OMPT support

See merge request dlb/dlb!2

* ompt_tool.c contains the OMPT start tool function and callbacks
* omp_thread_manager contains the LewI mechanism

* callbacks must modify omp_set_num_threads to make the API calls
    effective
* fix counter cpus_to_borrow increasing only when borrow succeeds
* when out of parallel regions, use DLB_SetMaxParallelism instead
* fix IntoBlockingCall by lending only non-current CPUs

Deprecate shmem_cpuinfo__thread_needs_rebinding and reimplement
shmem_cpuinfo__get_thread_binding.
Remove CPU binding from shmem_cpuinfo, now each binding is based on
thread_number taking the owned CPUs first.

PreInit function was not initializing values of cpuid.

* call DLB Init and Finalize if --ompt is enabled, no LeWI needed.
* call DLB Init before verbossing any OMPT message
* let DLB Init manage option inconsistencies
* reduce OMPT callback messages if all of them are ok

We want to distinguish the OMPT feature which intercepts
OpenMP routines from the automatic LeWI from OMPT callbacks.
Now, the option --ompt just enables OMPT interception and
the option --lewi-ompt accepts different values to construct
the policy.

This binary is not installed during the make install process.
It serves for debug purposes: runs until interrupted and
prints continuously the binding of each thread.

Set all CPU ids when the shmem is created, instead of doing it
only when a process registers.
This was causing erroneous bindings with unregistered CPUs.

* Remove generic shmem_cpuinfo__is_dirty, each CPU needs to find its own flag
* CPU binding is now tracked by guest, instead of by owner, this way simplifies
    the implementation of bindings for both cases DROM and LeWI
* Due to the previous point, thread_num must be invalidated when a process erases
    the guest field
* Update tests to reflect changes

* Init now does not need to compute the affinity mask
* Finalize must nullify some structures to not cause segfault
if called several times

Update OMPT related types to OpenMP TR6: 5.0 preview 2

Parallel tests
Add lit support for doing the 'make check' in parallel

See merge request dlb/dlb!1

Use a common header to generate a SHM_KEY unique to each test
The header also provides other shmem utilities like automatic
clean up on program finalization or abort.