blt_cuda_smoke compiled with default arch

I noticed that `blt_cuda_smoke` is built with default arch, while most of the code I want to test is built with extra arch flags.

That can lead to a `blt_cuda_smoke` that works when other exes do not.

This is a BLT usage thing we should make folks aware of.

Start by locating the BLT usage guidance for blt_cuda_smoke and the handling of extra architecture flags. Document that the smoke executable uses the default architecture while other executables may use additional flags, and explain the resulting testing limitation.

Written by the indexing model from the issue text.