Remove track_features for cuda builds?

[h-vetinari]

We've been weighing down the CUDA builds ever since they got added in #137.

Now that conda versions with __cuda are way more widespread, I think we could remove the track_features, and just rely on the (already existing!) run-dependence on __cuda, to ensure that people don't download the GPU bits senselessly.

Thoughts @conda-forge/arrow-cpp?

CC @conda-forge/core

[hmaarrfk]

I've found that the build number strategy is easier to understand in practice.

I'm in favor of removing track feature.