Remove CUDA whole compilation ODR violations

[robertmaynard]

Description

CUDA whole compilation mode requires that all kernels are only launched from TUs that compile them. Previously libcudf would compile a subset of kernels in separate TUs from where they are launched.
To keep compile times ( and library size ) as low as possible I have introduced a single C++ function call between the original call site and the kernel launch. In testing this neglibile differences on compile time and binary size.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[vyasr]

IIUC this is basically just adding an extra level of indirection so that the kernels are only compiled once and hidden beneath a single C++ launch function that can be called in different TUs? If so, LGTM.

[robertmaynard]

IIUC this is basically just adding an extra level of indirection so that the kernels are only compiled once and hidden beneath a single C++ launch function that can be called in different TUs? If so, LGTM.

That is correct

[robertmaynard]

/merge