Add functionality from the CUDA driver, not just the runtime API

[eyalroz]

The CUDA driver has all sorts of functionality not accessible directly, or at all, through the runtime API. Some of it is even very useful for implementing some of the proxy classes' methods (e.g. getting a single pointer attribute rather than all of them). Perhaps we should start adding them.

... this could theoretically be within #ifndef RUNTIME_API_ONLY.

[eyalroz]

An important piece of functionality to add - more stuff we can enqueue:

• Write a single value to a device memory location:

  CUresult cuStreamWriteValue32 ( CUstream stream, CUdeviceptr addr, cuuint32_t value, unsigned int  flags )
  CUresult cuStreamWriteValue64 ( CUstream stream, CUdeviceptr addr, cuuint64_t value, unsigned int  flags )
  

• Wait for a change in a device memory location (which could be mapped):

  CUresult cuStreamWaitValue32 ( CUstream stream, CUdeviceptr addr, cuuint32_t value, unsigned int  flags )
  CUresult cuStreamWaitValue64 ( CUstream stream, CUdeviceptr addr, cuuint64_t value, unsigned int  flags )
  

[eyalroz]

This could be done as part of separating this repository into 3 wrapper libraries: Runtime API, NVTX, and driver API (issue #176). But then - a lot of the driver API is redundant with the runtime API... :-(