NSight Compute: prevent API calls during precompilation

[Tim Besard]

Otherwise you'll end up attaching to the precompilation process.

Also, the docs are out of date, because it waits for the debugger now during using CUDA, and not when doing the first CUDA operation.

[Tim Besard]

This turns out to be tricky, as NSight Compute also breaks on version queries, which we do during platform augmentation and during forwards-compatible driver selection. So instead I've added to the docs that it's recommended to precompile all dependencies before profiling.