[FEA] Support JITting and Linking LTOIR from cuda source inputs

[isVoid]

Is your feature request related to a problem? Please describe.
After Pynvjitlink is introduced into Numba-CUDA, I would like to be able to jit cuda source files into LTOIR and link it with nvjitlink. This basically requires the linker lto flag to be respected by the jit and linker machinery.

Describe the solution you'd like
Add lto flag to nvrtc.compile, and add -dlto flag to compile option when set. In the Linker class, use add_ltoir in add_cu when self.lto is set.

[isVoid]

As a follow up to this PR - it would be helpful to support -ptx flag for nvjitlink to emit the PTX under LTOIR linkage, given NUMBA_DUMP_ASSEMBLY is set. This makes sure that Numba-CUDA shows the LTO-ed PTX and is more useful in debugging optimization related issues.