Stable version, options of DLB_ARGS have been added.

Fortran dlbf.h update

In order to honor the same behavior for all the functions in the API,
DLB_Lend must affect all the CPUs and it will no longer keep the
current CPU.

The function lewi_mask_DisableDLB was not properly triggering
the callbacks if a CPU needed to be disabled and enbaled by
other subprocess.

A race condition between two shmem finalization may occur if
one of them is made from a helper process that only opens the
shared memory to finalize other process. In that case both proceses
may think that they are the last one and both will try to delete
the shared memory. Ignore error in that case.

Even if process does not exist, it is much safer against race
conditions to lock the shared memory, look if the shmem is empty
and close it properly.

The process is now deregistered in shmem_cpuinfo from the policy
in order to manage the request queues.

Try to clean also shared memories of registered subprocesses.

* Do not increase counter of subprocesses_attached if
    initialization fails, and return the appropriate value
* Add field preregistered to detect if pid is so

Commit 0f9bd0e6 modified
the queue structure and the shmem cpuinfo version should
have been increased.

Allow hostname to be half of fmt_str length, at most.

Include also the requests queues and a more detailed legend.

- queue_proc_reqs_**pop** refactored to **get** since this
    function can remove elements in the middle
- Implement queue_proc_reqs_pop that always removes the
    front element
- Add some const modifiers
- Ensure on every remove operation that queue->tail points to
    a valid element, unless the queue is empty.

Working to fix a variation of MonitoringRegions

Fix #100

* Delay CPU_CLR if we are waiting for a bit in an assert_loop. It may cause race conditions
* Delay shmem_async_finalize after other shmem finalizations
* Initialize with --mode=async instead of using the test generator if there is only one flag