Add support for batched tasks and for CUDA-aware communications

[bosilca]

The idea for the batching is the following:

• tasks incarnations (aka. BODY) can be marked with the "batch" property allowing the runtime to provide the task with the entire list of ready tasks of the execution stream instead of just extracting the head.
• this list of ready tasks is in fact a ring, that can then be trimmed by the kernel and divided into batch and the rest. The rest of the tasks will be left in the ring, while the batch group will be submitted for execution.
• the kernel also needs to provide a callback into the gpu_task complete_stage, such that the runtime can call the specialized function able to complete all batched tasks.

The idea for the CUDA-aware communications is to use the task class data_affinity capability to locate the affinity of a task, and then allocate all incoming data for that task into the preferred location of that data. Thus, in order to take advantage of this capability, it is critical to hint the runtime where to locate data (parsec_advise_data_on_device). Note, that only the first successor is checked for the affinity.

The communication thread allocates data on a GPU by bypassing the GPU allocator, and hitting directly into the zone_malloc (which is now protected against concurrent access). Once tiles are allocated they will be incorporated into the GPU LRUs by the first call involving them, so the GPU becomes aware of this data rather quickly. The issue is in the other direction, as parsec has a tendency to keep the temporary data in the cache, it will never refill the zone_malloc, leading to starvation on device memory for the communication thread. This is, I think, the root cause for the sporadic deadlocks I encounter.

I have started to implement a different approach, where the GPU memory can be extracted from LRUs by the communication thread. Part of the code is in this PR, but I did not had the time to integrate it nicely.

Anyway, last but not least, a word about performance. On a small (4 A100 GPU, IB but no nvlink) I get
no CUDA-aware: [****] TIME(s) 7.53502 : dpotrf PxQxg= 2 2 1 NB= 1024 N= 65536 : 12452.141424 gflops - ENQ&PROG&DEST 8.28404 : 11326.255177 gflops - ENQ 0.49800 - DEST 0.25101
CUDA-aware: [****] TIME(s) 3.80899 : dpotrf PxQxg= 2 2 1 NB= 1024 N= 65536 : 24633.106084 gflops - ENQ&PROG&DEST 4.56897 : 20535.733216 gflops - ENQ 0.51097 - DEST 0.24902

[bosilca]

This is a leftover from another patch allowing typed locals. Will bring it in at a later date, meanwhile I will need to remove this code.

[bosilca]

For now this PR remains here until all parts have been merged. The code has been split in 4 unrelated parts:

• Add support for batched tasks. ICLDisco/parsec#668
• Fix use of calloc. ICLDisco/parsec#669
• Allow selection of a particular GPU (via the mask). ICLDisco/parsec#670
• Topic/cuda aware communications ICLDisco/parsec#671