Add wrappers for the CUDA graph API.

[maleadt]

Fixes #65.

I'm pretty happy with how the API turned out. See the tests for two examples, capture/instantiate/update/launch for full control, @captured wrapping all that. I didn't manage to outperform regular kernel launches though, but that's probably due to my limited benchmarking of the APIs.

Note that I started looking into this to implement a good mapslices implementation, but then I realized the function passed to mapslices allocates its output on every iteration, so we can't use the graph API for it.

[codecov]

Codecov Report

Merging #877 (71d5f39) into master (4549cb7) will decrease coverage by 0.02%.
The diff coverage is 74.19%.

@@            Coverage Diff             @@
##           master     #877      +/-   ##
==========================================
- Coverage   77.14%   77.11%   -0.03%     
==========================================
  Files         121      122       +1     
  Lines        7546     7608      +62     
==========================================
+ Hits         5821     5867      +46     
- Misses       1725     1741      +16     

Impacted Files	Coverage Δ
lib/cudadrv/graph.jl	74.19% <74.19%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 4549cb7...71d5f39. Read the comment docs.