Allow asynchronous building of gtscript stencils. (gt4py.cartesian)

[gronerl]

Fixes #83

This PR implements asynchronous building of stencils in cartesian gtscript (PR affects only gt4py.cartesian)

API:

• The eager keyword to the lazy_stencil decorator is extended. It can now be set to "async" which will schedule compilation immediately asynchronously. Wait is triggered by accessing LazyStencil.implementation (e.g. by means of __call__)
• There is also a wait_all() function in gt4py.cartesian.stencil_builder that is a barrier that can be used
• The maximum number of building processes is controlled through basically this:

build_settings: Dict[str, Any] = {
 # some more options omitted   
 "max_async_build_proc": os.environ.get("GT_MAX_ASYNC_BUILD_PROC", mp.cpu_count()),
}

ToDo:

• stress-test e.g. by means of use in pace
• update docs

Note that since StencilObject can not be pickled, it can't be returned from multiprocessing / result of concurrent.future with a ProcessPoolExecutor. In the implementation at hand, instead the load step is run multiple times (once in the building child process and once in the main process). If this is ok is to be seen (e.g. from `pace).

[gronerl]

Implements #83

[gronerl]

Since lazy_stencil is undocumented outside of the docstring, I just updated the docstring, without adding additional files.