Replace finalizer() with finalization queues

[muellren]

Off-heap memory allocated by DeviceArray objects are freed in the finalizer, i.e., cudaFree(...) is called in the objects finalize() method.

As finalize() is deprecated since Java 9, resource deallocation should be implemented using finalization queues.

The finalizer queue should be checked at certain points throughout the lifetime of the language context and memory buffers on the queue freed up with cudaFree(). The cleanup must be performed by the proper thread. Given grCUDA is registered as single-threaded language in Truffle, the deallocation must be performed by the main thread.

[muellren]

This does not only apply to device arrays but to other off-heap resources also.

• cuModule in CUDARuntime
• Handles in for cuML
• Handles for cuBLAS