object shredding, numerous refactorings

[lukasstadler]

This PR includes numerous (too many?) changes, mainly because they are somewhat interdependent. I can try to dice it up into several PRs if necessary.

• Replace all instances of "RuntimeException" with either GrCUDAException or GrCUDAInternalException, so that all thrown exceptions are proper TruffleException instances.
• Export the namespaces into the polyglot namespace (eval("grcuda","") will obtain the root namespace), this allows TAB completion in, e.g., Python.
• Add a "mockup" mode that doesn't load CUDA libraries and that can be used for testing basic DeviceArray access functionality.
• A skeleton for cuBLAS support (similar to cuML).
• A function to map from an existing array to a DeviceArray (MapDeviceArrayFunction). This function uses Truffle LoopNode to allow on stack replacement, i.e., fast execution even for the first call.
• Versions of DeviceArrayFunction and MapDeviceArrayFunction that are curried with the type (TypedDeviceArrayFunction, TypedMapDeviceArrayFunction).
• An initial version of the argument mapping and object shredding functionality.
• Some refactorings to simplify the code in CUDARuntime.
• Use Truffle options system to allow for "--grcuda.Xyz" style options.
• cuBLAS/cuML libraries are only loaded when the first function is called, and they are now enabled by default.
• Remove usage of deprecated "parse" API in CUDARuntime.

I'm looking forward to comments/feedback/...!

[muellren]

This PR provides support for #4 .

[lukasstadler]

I'm closing this, all changes were pulled out into separate PRs.