Cluster backend for shot parallelization

[hitomitak]

Summary

Enhancement of #1281

Enhance the execution option to enable shot parallelization.
When a user gives dask client as executor, Aer can execute a simulation by a shot on the distributed machines like HPC clusters.

When the noise simulator executes with the executor option, AerJobSet object copies Qobj or circuits and distributes them to the workers. After simulations, AerJobSet collects each result, sums up with each count, and combines them into one result object.

Example Usage:
Threadpool execution:

exc = ThreadPoolExecutor(max_workers=2)
qbackend = AerSimulator(noise_model=noise_model)
result = qbackend.run(circ_list, executor=exc).result()

Dask execution (Local only):

exc = Client(address=LocalCluster(n_workers=1, processes=True))
qbackend = AerSimulator(noise_model=noise_model)
result = qbackend.run(circ_list, executor=exc).result()

Dask execution (Multiple Nodes):
After setup dask environment in all nodes, set up a cluster in a master node as a following command.

% dask-ssh dask1 dask2 dask3 &

And simulation executes.

exc = Client('dask1:8786')
qbackend = AerSimulator(noise_model=noise_model)
result = qbackend.run(circ_list, executor=exc).result()

Details and comments

I executed QV (10 depth, 100 shots ) with a noise model and measured the execution time on 3VMs with Dask executor.
Env: 3VMs, OS: CentOS8.0, Memory 64 GB, CPU: 32 x 2.0 GHz or higher Cores.
1 Node : a normal Aer simulator
1 Node (DASK) : only one node with dask executor
2 Node (DASK) : two nodes with dask executor
3 Node (DASK) : three nodes with dask executor