Fixed and improved custom OpenCL memory buffer wrapping.

Now vectors can also deal with offset and strides within the provided buffer.