Parallelize the installation script

[xuzhao9]

Installing package dependencies should be independent, and we should install them in parallel.

[FindHao]

I suggest making it configurable. Maybe no need to expose an API to the upper install.py but leave a variable or argument to disable it. Because I found an error named device busy caused by model installation in some clusters using NFS file system. I guess it is because the installation has too many write operations, and the cluster has a too harsh policy on write operations. After manually inserting a time.sleep(1) for each model, the installation becomes stable without any error. It happens several times when I use that cluster.

[xuzhao9]

We could limit the parallelism with a configurable process pool, for example, by default we allow maximum 5 processes running for installation. If we set it to 1, the behaviour should be the same as the current approach.