Cargo features for CUDA version (updated Dec. 10)

[peterhj]

This commit adds feature flags in Cargo.toml for specific installed versions of the CUDA toolkit. The design is basically the same as that used by clang-sys (https://github.com/KyleMayes/clang-sys/blob/master/Cargo.toml).

Edit (Dec. 10): Assuming the current static bindings for CUDA 8.0, I added static API version checks (using the static_assertions crate) and conditional bindings to the FP16 API (currently just an empty module). This way in which the version features are used should be independent of whether static or dynamic bindings are used in cuda-sys.

I also set the corresponding cargo features for each nvidia docker image in the CircleCI config. (Because cuda-sys has static bindings to CUDA 8.0, I set the default feature to be "cuda_8_0" in Cargo.toml, but also set --no-default-features in the CI runs.)

These feature flags are not currently used by the library or build script yet to actually select the CUDA version (e.g. for dynamic linking), but that could be done in another PR (pending discussion in #4).

[kngwyu]

What does gte_ prefix mean?

And I think we should add a version flag after actually testing it.

[AndrewGaspar]

What does gte_ prefix mean?

I assume it means "Greater-than-or-equal". I.e. the feature indicates at least that minimum version is available.

[peterhj]

What does gte_ prefix mean?

The gte_ (Greater Than or Equal) features can be used in the library itself, e.g. if certain APIs needed wrapping which differed between versions. The features without gte_ are specified in Cargo.toml by users of cuda-sys.

For example, one difference in CUDA APIs between versions is the addition of f16 support somewhere around 8.0 or 9.0, which added a new header among other changes.

[peterhj]

I added some minimal usage of the new API version features (just for 8.0, since that's what cuda-sys binds to currently). I also updated the CI config to set the right features in the environment (more details in the top post).