[fix] provide patch to add priority to task without support frm the compiler, such as ICC 2017 ou 2018

OMP_MAX_TASK_PRIORITY defines the number of dynamically created queue

- queue can be integrated to aggregation protocol
- framework is in place to capture graph construction in order to reschedule it.
- kmp_taskreschedule.cpp/h define markers for graph construction and heuristic
to reschedule list of ready tasks in order to improve locality

In order to active double linked list task: configure library with -DLIBOMP_USE_LINKED_DEQUEUE=true
In order to active rescheduling of the ready task list; configure library with -DLIBOMP_USE_REORDER4LOCALITY=true

Fix a huge issue in reordering by checking if a splice will fail (bad complexity)
Author: Thierry Gautier and Jérôme Richard

Each task has 2 keys to store task specific values.
These values are recorded into the trace.
csv extraction add two column Key0 and Key1 in tasks.csv

Thread to collect counter with kid=-1. Change indexing in fd table for the trace recorder
- change the initialization order between hws/topo and uncore that requires it
- uncore event not rely on parallel region definition
- larger name for perf counter
- generation in counters.cscv

These extensions are defined to constraint the scheduling with constrainted memory resources.
The idea is to let the user to define memory attributs for its OpenMP task and to set memory limits.
The scheduler will try to do the best by memory usage under the memory limits.

Attribut is given by calling new OpenMP runtime function.
  - omp_set_task_alloc_size(rsrc,size,ptr): to give to the runtime information on the usage of memory resource for the next created task.
  - omp_set_task_free_size(rsrc,size,ptr): to give to the runtime information information on task that may potential reduce memory by size bytes.
  - omp_info_alloc_memory(rsrc, size, ptr): used to indicate that 'size' byte on the resource 'rsrc' is allocated.
  - omp_info_free_memory(rsrc, size, ptr): used to indicate that 'size' byte on the resource 'rsrc' is freed.
  - omp_set_memory_limit( rsrc, size_max, percent ): to set limits on the set of resources in rsrc (xored value).
  The percent (float) will be used for future usage.
  By default there is no limit for each kind resource, nevertheless the defined limit has default
  value given by env. var. KAAPI_SCHED_MEM_LIMIT before any call to omp_set_memory_limit.

Add counter of used memory from declaration of the task' allocation and the free call in the program.
Katracereader generates memory.csv

[add] schedinfo event to the trace. The event records cpu/node of the running thread. Generated before task execution during steal operation or at the begin and the end of parallel region.

[add] schedinfo event to the trace. The event records cpu/node of the running thread. Generated before task execution during steal operation or at the begin and the end of parallel region.

 - fix native cpuset size
 - change default topology method (default to hwloc)

[add] support for omp_task_declare_dependencies_array and omp_task_declare_dependencies_array_noalias.
These OMP runtime functions let the user to pass an array of count items, each item of size 'extent', to the runtime
in order to indicate that there are dependencies on each item of the array.

- correct bug with gcc entry point

Conflicts:
	runtime/src/kmp.h
	runtime/src/kmp_taskdeps.cpp