Config struct + reductions

[brabreda]

This pull request introduces:

• a config object that can be used to use specific features depending on a backend.
• reduction API

Another idea would be to include ndrange in this config object. With this, we would be able to base the kernellaunch on this object so no need to define groupsize and ndrange again.

[vchuravy]

Thank you, this is interesting. A couple quick comments:

1. The dependency relationship between GPUArrays and KernelAbstractions should be inverted. E.g. KA should be a dependency of GPUArrays and used there to implement map reduce.
2. The new functionality requires tests and documentation.
3. How does Config differ from Kernel? Could we merge them into the same steuct?

[brabreda]

I will fix the first 2 points.

For me, config is rather a way of defining an algorithm rather than the kernel. For example max_ndrange does not represent the maximum amount of threads that the GPU supports. It just tells that the shouldn't be more thread as it may impact the performance for that operation.