allow f suffix in CUDA compute capabilities

[bedroge]

CUDA compute capabilities can now also have an f suffix, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#feature-set-compiler-targets.

Note that it's also changed in the regex that is used to extract the CUDA architectures from binaries. I wasn't sure if the f suffix would show up there as well, but I tested that by compiling something manually (with -arch compute_100f) and running cuobjdump on it:

Fatbin ptx code:
================
arch = sm_100
code version = [8,8]
host = linux
compile_size = 64bit
compressed
ptxasOptions = 

Fatbin ptx code:
================
arch = sm_100f
code version = [8,8]
host = linux
compile_size = 64bit
compressed
ptxasOptions = 

So, it does indeed have sm_100f.

[boegel]

lgtm

[boegel]

lgtm