Re-add convolution kernel

I have looked at the `convolution!` code but there are some details I am missing.

- Can the kernel have any shape?
- It seems the method returns the same shape of the original but there is no option for padding or not padding.

Wouldn't it be nice to have some documentation and several tests showing how to use it?

I have done some tests:
https://github.com/davidbp/learn_julia/blob/master/GPU_compute/GPUArrays/GPUArrays_convolution.ipynb

Start with the `convolution!` implementation and the linked GPUArrays convolution notebook. Use the notebook's experiments to determine the supported kernel shapes and padding behavior, then document those behaviors and add tests covering the examples. Done means the usage and shape behavior are documented and the tests pass.

Written by the indexing model from the issue text.