fix: Strip common arch-specific flags before processing

[varungandhi-src]

Generalizes the argument dropping mechanism used specifically
for nvcc to also work for Clang and GCC, and skips some
known arch-specific flags for Clang and GCC, to reduce risk
of unknown flags hit in practice on other architectures.

Fixes GRAPH-1187. See comments for more context.

I've not added a full-blown mechanism to skip arbitrary flags
(e.g. via an extra flag), but we could do that later if necessary.

Added some basic unit tests at the level of the command-line
tweaking. It's a bit more difficult to test this logic at a higher-level,
because the layer above it does I/O for compiler detection.

[varungandhi-src]

Merging now so that we can cut a release.