[FEA] Update cmake to only build for present GPU

[jrhemstad]

Is your feature request related to a problem? Please describe.

As of 388dd12 libcudf's full build compile time using ninja is: 7m 40s

The longer the compile time, the slower the development process. As such, we should do what we can to reduce compile time. Currently, libcudf's cmake configuration builds a fatbin for both sm60 and sm70.

If I disable the fatbin and only generate PTX for my target architecture (compute_70), compile time is: 3m 38s

However, this comes at a significant runtime cost: as soon as you try and run any tests, the CUDA runtime will JIT compile all of the library's PTX into SASS (which takes a very long time).

Instead, if we generate a fatbin for only a single GPU (e.g., compute_70,sm_70), compile time is: 4m 54s

Describe the solution you'd like

Update the CMake build to only generate a fatbin for the architecture present on the machine.

Additional context
Related: #1099 #581

[jrhemstad]

cuML looks to already have much of the machinery for detecting the target arch: https://github.com/rapidsai/cuml/blob/branch-0.10/cpp/cmake/EvalGpuArchs.cmake

[jrhemstad]

Closed by #2955